]>
Commit | Line | Data |
---|---|---|
34dc7c2f BB |
1 | /* |
2 | * CDDL HEADER START | |
3 | * | |
4 | * The contents of this file are subject to the terms of the | |
5 | * Common Development and Distribution License (the "License"). | |
6 | * You may not use this file except in compliance with the License. | |
7 | * | |
8 | * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE | |
9 | * or http://www.opensolaris.org/os/licensing. | |
10 | * See the License for the specific language governing permissions | |
11 | * and limitations under the License. | |
12 | * | |
13 | * When distributing Covered Code, include this CDDL HEADER in each | |
14 | * file and include the License file at usr/src/OPENSOLARIS.LICENSE. | |
15 | * If applicable, add the following below this CDDL HEADER, with the | |
16 | * fields enclosed by brackets "[]" replaced with your own identifying | |
17 | * information: Portions Copyright [yyyy] [name of copyright owner] | |
18 | * | |
19 | * CDDL HEADER END | |
20 | */ | |
21 | /* | |
428870ff | 22 | * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved. |
492f64e9 | 23 | * Copyright (c) 2011, 2018 by Delphix. All rights reserved. |
a38718a6 | 24 | * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved. |
cc99f275 | 25 | * Copyright (c) 2017, Intel Corporation. |
34dc7c2f BB |
26 | */ |
27 | ||
f1512ee6 | 28 | #include <sys/sysmacros.h> |
34dc7c2f BB |
29 | #include <sys/zfs_context.h> |
30 | #include <sys/fm/fs/zfs.h> | |
31 | #include <sys/spa.h> | |
32 | #include <sys/txg.h> | |
33 | #include <sys/spa_impl.h> | |
34 | #include <sys/vdev_impl.h> | |
35 | #include <sys/zio_impl.h> | |
36 | #include <sys/zio_compress.h> | |
37 | #include <sys/zio_checksum.h> | |
428870ff BB |
38 | #include <sys/dmu_objset.h> |
39 | #include <sys/arc.h> | |
40 | #include <sys/ddt.h> | |
9b67f605 | 41 | #include <sys/blkptr.h> |
b0bc7a84 | 42 | #include <sys/zfeature.h> |
d4a72f23 | 43 | #include <sys/dsl_scan.h> |
3dfb57a3 | 44 | #include <sys/metaslab_impl.h> |
193a37cb | 45 | #include <sys/time.h> |
26ef0cc7 | 46 | #include <sys/trace_zio.h> |
a6255b7f | 47 | #include <sys/abd.h> |
b5256303 | 48 | #include <sys/dsl_crypt.h> |
492f64e9 | 49 | #include <sys/cityhash.h> |
34dc7c2f | 50 | |
34dc7c2f BB |
51 | /* |
52 | * ========================================================================== | |
53 | * I/O type descriptions | |
54 | * ========================================================================== | |
55 | */ | |
e8b96c60 | 56 | const char *zio_type_name[ZIO_TYPES] = { |
3dfb57a3 DB |
57 | /* |
58 | * Note: Linux kernel thread name length is limited | |
59 | * so these names will differ from upstream open zfs. | |
60 | */ | |
451041db | 61 | "z_null", "z_rd", "z_wr", "z_fr", "z_cl", "z_ioctl" |
428870ff | 62 | }; |
34dc7c2f | 63 | |
27f2b90d | 64 | int zio_dva_throttle_enabled = B_TRUE; |
3dfb57a3 | 65 | |
34dc7c2f BB |
66 | /* |
67 | * ========================================================================== | |
68 | * I/O kmem caches | |
69 | * ========================================================================== | |
70 | */ | |
71 | kmem_cache_t *zio_cache; | |
d164b209 | 72 | kmem_cache_t *zio_link_cache; |
34dc7c2f BB |
73 | kmem_cache_t *zio_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; |
74 | kmem_cache_t *zio_data_buf_cache[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
a6255b7f DQ |
75 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
76 | uint64_t zio_buf_cache_allocs[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
77 | uint64_t zio_buf_cache_frees[SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT]; | |
78 | #endif | |
79 | ||
ad796b8a TH |
80 | /* Mark IOs as "slow" if they take longer than 30 seconds */ |
81 | int zio_slow_io_ms = (30 * MILLISEC); | |
34dc7c2f | 82 | |
fcff0f35 PD |
83 | #define BP_SPANB(indblkshift, level) \ |
84 | (((uint64_t)1) << ((level) * ((indblkshift) - SPA_BLKPTRSHIFT))) | |
85 | #define COMPARE_META_LEVEL 0x80000000ul | |
55d85d5a GW |
86 | /* |
87 | * The following actions directly effect the spa's sync-to-convergence logic. | |
88 | * The values below define the sync pass when we start performing the action. | |
89 | * Care should be taken when changing these values as they directly impact | |
90 | * spa_sync() performance. Tuning these values may introduce subtle performance | |
91 | * pathologies and should only be done in the context of performance analysis. | |
92 | * These tunables will eventually be removed and replaced with #defines once | |
93 | * enough analysis has been done to determine optimal values. | |
94 | * | |
95 | * The 'zfs_sync_pass_deferred_free' pass must be greater than 1 to ensure that | |
96 | * regular blocks are not deferred. | |
97 | */ | |
98 | int zfs_sync_pass_deferred_free = 2; /* defer frees starting in this pass */ | |
99 | int zfs_sync_pass_dont_compress = 5; /* don't compress starting in this pass */ | |
100 | int zfs_sync_pass_rewrite = 2; /* rewrite new bps starting in this pass */ | |
101 | ||
34dc7c2f | 102 | /* |
b128c09f BB |
103 | * An allocating zio is one that either currently has the DVA allocate |
104 | * stage set or will have it later in its lifetime. | |
34dc7c2f | 105 | */ |
428870ff BB |
106 | #define IO_IS_ALLOCATING(zio) ((zio)->io_orig_pipeline & ZIO_STAGE_DVA_ALLOCATE) |
107 | ||
c409e464 | 108 | int zio_requeue_io_start_cut_in_line = 1; |
428870ff BB |
109 | |
110 | #ifdef ZFS_DEBUG | |
111 | int zio_buf_debug_limit = 16384; | |
112 | #else | |
113 | int zio_buf_debug_limit = 0; | |
114 | #endif | |
34dc7c2f | 115 | |
da6b4005 NB |
116 | static inline void __zio_execute(zio_t *zio); |
117 | ||
3dfb57a3 DB |
118 | static void zio_taskq_dispatch(zio_t *, zio_taskq_type_t, boolean_t); |
119 | ||
34dc7c2f BB |
120 | void |
121 | zio_init(void) | |
122 | { | |
123 | size_t c; | |
124 | vmem_t *data_alloc_arena = NULL; | |
125 | ||
3941503c BB |
126 | zio_cache = kmem_cache_create("zio_cache", |
127 | sizeof (zio_t), 0, NULL, NULL, NULL, NULL, NULL, 0); | |
d164b209 | 128 | zio_link_cache = kmem_cache_create("zio_link_cache", |
6795a698 | 129 | sizeof (zio_link_t), 0, NULL, NULL, NULL, NULL, NULL, 0); |
34dc7c2f BB |
130 | |
131 | /* | |
132 | * For small buffers, we want a cache for each multiple of | |
f1512ee6 MA |
133 | * SPA_MINBLOCKSIZE. For larger buffers, we want a cache |
134 | * for each quarter-power of 2. | |
34dc7c2f BB |
135 | */ |
136 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
137 | size_t size = (c + 1) << SPA_MINBLOCKSHIFT; | |
138 | size_t p2 = size; | |
139 | size_t align = 0; | |
6442f3cf | 140 | size_t cflags = (size > zio_buf_debug_limit) ? KMC_NODEBUG : 0; |
34dc7c2f | 141 | |
34328f3c | 142 | #if defined(_ILP32) && defined(_KERNEL) |
f1512ee6 MA |
143 | /* |
144 | * Cache size limited to 1M on 32-bit platforms until ARC | |
145 | * buffers no longer require virtual address space. | |
146 | */ | |
147 | if (size > zfs_max_recordsize) | |
148 | break; | |
149 | #endif | |
150 | ||
151 | while (!ISP2(p2)) | |
34dc7c2f BB |
152 | p2 &= p2 - 1; |
153 | ||
498877ba MA |
154 | #ifndef _KERNEL |
155 | /* | |
156 | * If we are using watchpoints, put each buffer on its own page, | |
157 | * to eliminate the performance overhead of trapping to the | |
158 | * kernel when modifying a non-watched buffer that shares the | |
159 | * page with a watched buffer. | |
160 | */ | |
161 | if (arc_watch && !IS_P2ALIGNED(size, PAGESIZE)) | |
162 | continue; | |
fcf64f45 BB |
163 | /* |
164 | * Here's the problem - on 4K native devices in userland on | |
165 | * Linux using O_DIRECT, buffers must be 4K aligned or I/O | |
166 | * will fail with EINVAL, causing zdb (and others) to coredump. | |
167 | * Since userland probably doesn't need optimized buffer caches, | |
168 | * we just force 4K alignment on everything. | |
169 | */ | |
170 | align = 8 * SPA_MINBLOCKSIZE; | |
171 | #else | |
24fa2034 | 172 | if (size < PAGESIZE) { |
34dc7c2f | 173 | align = SPA_MINBLOCKSIZE; |
498877ba | 174 | } else if (IS_P2ALIGNED(size, p2 >> 2)) { |
24fa2034 | 175 | align = PAGESIZE; |
34dc7c2f | 176 | } |
fcf64f45 | 177 | #endif |
34dc7c2f BB |
178 | |
179 | if (align != 0) { | |
180 | char name[36]; | |
181 | (void) sprintf(name, "zio_buf_%lu", (ulong_t)size); | |
182 | zio_buf_cache[c] = kmem_cache_create(name, size, | |
6442f3cf | 183 | align, NULL, NULL, NULL, NULL, NULL, cflags); |
34dc7c2f BB |
184 | |
185 | (void) sprintf(name, "zio_data_buf_%lu", (ulong_t)size); | |
186 | zio_data_buf_cache[c] = kmem_cache_create(name, size, | |
ae6ba3db | 187 | align, NULL, NULL, NULL, NULL, |
6442f3cf | 188 | data_alloc_arena, cflags); |
34dc7c2f BB |
189 | } |
190 | } | |
191 | ||
192 | while (--c != 0) { | |
193 | ASSERT(zio_buf_cache[c] != NULL); | |
194 | if (zio_buf_cache[c - 1] == NULL) | |
195 | zio_buf_cache[c - 1] = zio_buf_cache[c]; | |
196 | ||
197 | ASSERT(zio_data_buf_cache[c] != NULL); | |
198 | if (zio_data_buf_cache[c - 1] == NULL) | |
199 | zio_data_buf_cache[c - 1] = zio_data_buf_cache[c]; | |
200 | } | |
201 | ||
34dc7c2f | 202 | zio_inject_init(); |
9759c60f ED |
203 | |
204 | lz4_init(); | |
34dc7c2f BB |
205 | } |
206 | ||
207 | void | |
208 | zio_fini(void) | |
209 | { | |
210 | size_t c; | |
211 | kmem_cache_t *last_cache = NULL; | |
212 | kmem_cache_t *last_data_cache = NULL; | |
213 | ||
214 | for (c = 0; c < SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT; c++) { | |
f1512ee6 MA |
215 | #ifdef _ILP32 |
216 | /* | |
217 | * Cache size limited to 1M on 32-bit platforms until ARC | |
218 | * buffers no longer require virtual address space. | |
219 | */ | |
220 | if (((c + 1) << SPA_MINBLOCKSHIFT) > zfs_max_recordsize) | |
221 | break; | |
a6255b7f DQ |
222 | #endif |
223 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) | |
224 | if (zio_buf_cache_allocs[c] != zio_buf_cache_frees[c]) | |
225 | (void) printf("zio_fini: [%d] %llu != %llu\n", | |
226 | (int)((c + 1) << SPA_MINBLOCKSHIFT), | |
227 | (long long unsigned)zio_buf_cache_allocs[c], | |
228 | (long long unsigned)zio_buf_cache_frees[c]); | |
f1512ee6 | 229 | #endif |
34dc7c2f BB |
230 | if (zio_buf_cache[c] != last_cache) { |
231 | last_cache = zio_buf_cache[c]; | |
232 | kmem_cache_destroy(zio_buf_cache[c]); | |
233 | } | |
234 | zio_buf_cache[c] = NULL; | |
235 | ||
236 | if (zio_data_buf_cache[c] != last_data_cache) { | |
237 | last_data_cache = zio_data_buf_cache[c]; | |
238 | kmem_cache_destroy(zio_data_buf_cache[c]); | |
239 | } | |
240 | zio_data_buf_cache[c] = NULL; | |
241 | } | |
242 | ||
d164b209 | 243 | kmem_cache_destroy(zio_link_cache); |
34dc7c2f BB |
244 | kmem_cache_destroy(zio_cache); |
245 | ||
246 | zio_inject_fini(); | |
9759c60f ED |
247 | |
248 | lz4_fini(); | |
34dc7c2f BB |
249 | } |
250 | ||
251 | /* | |
252 | * ========================================================================== | |
253 | * Allocate and free I/O buffers | |
254 | * ========================================================================== | |
255 | */ | |
256 | ||
257 | /* | |
258 | * Use zio_buf_alloc to allocate ZFS metadata. This data will appear in a | |
259 | * crashdump if the kernel panics, so use it judiciously. Obviously, it's | |
260 | * useful to inspect ZFS metadata, but if possible, we should avoid keeping | |
261 | * excess / transient data in-core during a crashdump. | |
262 | */ | |
263 | void * | |
264 | zio_buf_alloc(size_t size) | |
265 | { | |
266 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
267 | ||
63e3a861 | 268 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
269 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
270 | atomic_add_64(&zio_buf_cache_allocs[c], 1); | |
271 | #endif | |
34dc7c2f | 272 | |
efcd79a8 | 273 | return (kmem_cache_alloc(zio_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
274 | } |
275 | ||
276 | /* | |
277 | * Use zio_data_buf_alloc to allocate data. The data will not appear in a | |
278 | * crashdump if the kernel panics. This exists so that we will limit the amount | |
279 | * of ZFS data that shows up in a kernel crashdump. (Thus reducing the amount | |
280 | * of kernel heap dumped to disk when the kernel panics) | |
281 | */ | |
282 | void * | |
283 | zio_data_buf_alloc(size_t size) | |
284 | { | |
285 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
286 | ||
63e3a861 | 287 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f | 288 | |
efcd79a8 | 289 | return (kmem_cache_alloc(zio_data_buf_cache[c], KM_PUSHPAGE)); |
34dc7c2f BB |
290 | } |
291 | ||
292 | void | |
293 | zio_buf_free(void *buf, size_t size) | |
294 | { | |
295 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
296 | ||
63e3a861 | 297 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
a6255b7f DQ |
298 | #if defined(ZFS_DEBUG) && !defined(_KERNEL) |
299 | atomic_add_64(&zio_buf_cache_frees[c], 1); | |
300 | #endif | |
34dc7c2f BB |
301 | |
302 | kmem_cache_free(zio_buf_cache[c], buf); | |
303 | } | |
304 | ||
305 | void | |
306 | zio_data_buf_free(void *buf, size_t size) | |
307 | { | |
308 | size_t c = (size - 1) >> SPA_MINBLOCKSHIFT; | |
309 | ||
63e3a861 | 310 | VERIFY3U(c, <, SPA_MAXBLOCKSIZE >> SPA_MINBLOCKSHIFT); |
34dc7c2f BB |
311 | |
312 | kmem_cache_free(zio_data_buf_cache[c], buf); | |
313 | } | |
314 | ||
84c07ada GN |
315 | static void |
316 | zio_abd_free(void *abd, size_t size) | |
317 | { | |
318 | abd_free((abd_t *)abd); | |
319 | } | |
320 | ||
34dc7c2f BB |
321 | /* |
322 | * ========================================================================== | |
323 | * Push and pop I/O transform buffers | |
324 | * ========================================================================== | |
325 | */ | |
d3c2ae1c | 326 | void |
a6255b7f | 327 | zio_push_transform(zio_t *zio, abd_t *data, uint64_t size, uint64_t bufsize, |
e9aa730c | 328 | zio_transform_func_t *transform) |
34dc7c2f | 329 | { |
79c76d5b | 330 | zio_transform_t *zt = kmem_alloc(sizeof (zio_transform_t), KM_SLEEP); |
34dc7c2f | 331 | |
a6255b7f DQ |
332 | /* |
333 | * Ensure that anyone expecting this zio to contain a linear ABD isn't | |
334 | * going to get a nasty surprise when they try to access the data. | |
335 | */ | |
336 | IMPLY(abd_is_linear(zio->io_abd), abd_is_linear(data)); | |
337 | ||
338 | zt->zt_orig_abd = zio->io_abd; | |
b128c09f | 339 | zt->zt_orig_size = zio->io_size; |
34dc7c2f | 340 | zt->zt_bufsize = bufsize; |
b128c09f | 341 | zt->zt_transform = transform; |
34dc7c2f BB |
342 | |
343 | zt->zt_next = zio->io_transform_stack; | |
344 | zio->io_transform_stack = zt; | |
345 | ||
a6255b7f | 346 | zio->io_abd = data; |
34dc7c2f BB |
347 | zio->io_size = size; |
348 | } | |
349 | ||
d3c2ae1c | 350 | void |
b128c09f | 351 | zio_pop_transforms(zio_t *zio) |
34dc7c2f | 352 | { |
b128c09f BB |
353 | zio_transform_t *zt; |
354 | ||
355 | while ((zt = zio->io_transform_stack) != NULL) { | |
356 | if (zt->zt_transform != NULL) | |
357 | zt->zt_transform(zio, | |
a6255b7f | 358 | zt->zt_orig_abd, zt->zt_orig_size); |
34dc7c2f | 359 | |
428870ff | 360 | if (zt->zt_bufsize != 0) |
a6255b7f | 361 | abd_free(zio->io_abd); |
34dc7c2f | 362 | |
a6255b7f | 363 | zio->io_abd = zt->zt_orig_abd; |
b128c09f BB |
364 | zio->io_size = zt->zt_orig_size; |
365 | zio->io_transform_stack = zt->zt_next; | |
34dc7c2f | 366 | |
b128c09f | 367 | kmem_free(zt, sizeof (zio_transform_t)); |
34dc7c2f BB |
368 | } |
369 | } | |
370 | ||
b128c09f BB |
371 | /* |
372 | * ========================================================================== | |
b5256303 | 373 | * I/O transform callbacks for subblocks, decompression, and decryption |
b128c09f BB |
374 | * ========================================================================== |
375 | */ | |
376 | static void | |
a6255b7f | 377 | zio_subblock(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f BB |
378 | { |
379 | ASSERT(zio->io_size > size); | |
380 | ||
381 | if (zio->io_type == ZIO_TYPE_READ) | |
a6255b7f | 382 | abd_copy(data, zio->io_abd, size); |
b128c09f BB |
383 | } |
384 | ||
385 | static void | |
a6255b7f | 386 | zio_decompress(zio_t *zio, abd_t *data, uint64_t size) |
b128c09f | 387 | { |
a6255b7f DQ |
388 | if (zio->io_error == 0) { |
389 | void *tmp = abd_borrow_buf(data, size); | |
390 | int ret = zio_decompress_data(BP_GET_COMPRESS(zio->io_bp), | |
391 | zio->io_abd, tmp, zio->io_size, size); | |
392 | abd_return_buf_copy(data, tmp, size); | |
393 | ||
c3bd3fb4 TC |
394 | if (zio_injection_enabled && ret == 0) |
395 | ret = zio_handle_fault_injection(zio, EINVAL); | |
396 | ||
a6255b7f DQ |
397 | if (ret != 0) |
398 | zio->io_error = SET_ERROR(EIO); | |
399 | } | |
b128c09f BB |
400 | } |
401 | ||
b5256303 TC |
402 | static void |
403 | zio_decrypt(zio_t *zio, abd_t *data, uint64_t size) | |
404 | { | |
405 | int ret; | |
406 | void *tmp; | |
407 | blkptr_t *bp = zio->io_bp; | |
ae76f45c TC |
408 | spa_t *spa = zio->io_spa; |
409 | uint64_t dsobj = zio->io_bookmark.zb_objset; | |
b5256303 TC |
410 | uint64_t lsize = BP_GET_LSIZE(bp); |
411 | dmu_object_type_t ot = BP_GET_TYPE(bp); | |
412 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
413 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
414 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
415 | boolean_t no_crypt = B_FALSE; | |
416 | ||
417 | ASSERT(BP_USES_CRYPT(bp)); | |
418 | ASSERT3U(size, !=, 0); | |
419 | ||
420 | if (zio->io_error != 0) | |
421 | return; | |
422 | ||
423 | /* | |
424 | * Verify the cksum of MACs stored in an indirect bp. It will always | |
425 | * be possible to verify this since it does not require an encryption | |
426 | * key. | |
427 | */ | |
428 | if (BP_HAS_INDIRECT_MAC_CKSUM(bp)) { | |
429 | zio_crypt_decode_mac_bp(bp, mac); | |
430 | ||
431 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF) { | |
432 | /* | |
433 | * We haven't decompressed the data yet, but | |
434 | * zio_crypt_do_indirect_mac_checksum() requires | |
435 | * decompressed data to be able to parse out the MACs | |
436 | * from the indirect block. We decompress it now and | |
437 | * throw away the result after we are finished. | |
438 | */ | |
439 | tmp = zio_buf_alloc(lsize); | |
440 | ret = zio_decompress_data(BP_GET_COMPRESS(bp), | |
441 | zio->io_abd, tmp, zio->io_size, lsize); | |
442 | if (ret != 0) { | |
443 | ret = SET_ERROR(EIO); | |
444 | goto error; | |
445 | } | |
446 | ret = zio_crypt_do_indirect_mac_checksum(B_FALSE, | |
447 | tmp, lsize, BP_SHOULD_BYTESWAP(bp), mac); | |
448 | zio_buf_free(tmp, lsize); | |
449 | } else { | |
450 | ret = zio_crypt_do_indirect_mac_checksum_abd(B_FALSE, | |
451 | zio->io_abd, size, BP_SHOULD_BYTESWAP(bp), mac); | |
452 | } | |
453 | abd_copy(data, zio->io_abd, size); | |
454 | ||
be9a5c35 TC |
455 | if (zio_injection_enabled && ot != DMU_OT_DNODE && ret == 0) { |
456 | ret = zio_handle_decrypt_injection(spa, | |
457 | &zio->io_bookmark, ot, ECKSUM); | |
458 | } | |
b5256303 TC |
459 | if (ret != 0) |
460 | goto error; | |
461 | ||
462 | return; | |
463 | } | |
464 | ||
465 | /* | |
466 | * If this is an authenticated block, just check the MAC. It would be | |
467 | * nice to separate this out into its own flag, but for the moment | |
468 | * enum zio_flag is out of bits. | |
469 | */ | |
470 | if (BP_IS_AUTHENTICATED(bp)) { | |
471 | if (ot == DMU_OT_OBJSET) { | |
ae76f45c TC |
472 | ret = spa_do_crypt_objset_mac_abd(B_FALSE, spa, |
473 | dsobj, zio->io_abd, size, BP_SHOULD_BYTESWAP(bp)); | |
b5256303 TC |
474 | } else { |
475 | zio_crypt_decode_mac_bp(bp, mac); | |
ae76f45c TC |
476 | ret = spa_do_crypt_mac_abd(B_FALSE, spa, dsobj, |
477 | zio->io_abd, size, mac); | |
be9a5c35 TC |
478 | if (zio_injection_enabled && ret == 0) { |
479 | ret = zio_handle_decrypt_injection(spa, | |
480 | &zio->io_bookmark, ot, ECKSUM); | |
481 | } | |
b5256303 TC |
482 | } |
483 | abd_copy(data, zio->io_abd, size); | |
484 | ||
485 | if (ret != 0) | |
486 | goto error; | |
487 | ||
488 | return; | |
489 | } | |
490 | ||
491 | zio_crypt_decode_params_bp(bp, salt, iv); | |
492 | ||
493 | if (ot == DMU_OT_INTENT_LOG) { | |
494 | tmp = abd_borrow_buf_copy(zio->io_abd, sizeof (zil_chain_t)); | |
495 | zio_crypt_decode_mac_zil(tmp, mac); | |
496 | abd_return_buf(zio->io_abd, tmp, sizeof (zil_chain_t)); | |
497 | } else { | |
498 | zio_crypt_decode_mac_bp(bp, mac); | |
499 | } | |
500 | ||
be9a5c35 TC |
501 | ret = spa_do_crypt_abd(B_FALSE, spa, &zio->io_bookmark, BP_GET_TYPE(bp), |
502 | BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), salt, iv, mac, size, data, | |
503 | zio->io_abd, &no_crypt); | |
b5256303 TC |
504 | if (no_crypt) |
505 | abd_copy(data, zio->io_abd, size); | |
506 | ||
507 | if (ret != 0) | |
508 | goto error; | |
509 | ||
510 | return; | |
511 | ||
512 | error: | |
513 | /* assert that the key was found unless this was speculative */ | |
be9a5c35 | 514 | ASSERT(ret != EACCES || (zio->io_flags & ZIO_FLAG_SPECULATIVE)); |
b5256303 TC |
515 | |
516 | /* | |
517 | * If there was a decryption / authentication error return EIO as | |
518 | * the io_error. If this was not a speculative zio, create an ereport. | |
519 | */ | |
520 | if (ret == ECKSUM) { | |
a2c2ed1b | 521 | zio->io_error = SET_ERROR(EIO); |
b5256303 | 522 | if ((zio->io_flags & ZIO_FLAG_SPECULATIVE) == 0) { |
be9a5c35 | 523 | spa_log_error(spa, &zio->io_bookmark); |
b5256303 | 524 | zfs_ereport_post(FM_EREPORT_ZFS_AUTHENTICATION, |
ae76f45c | 525 | spa, NULL, &zio->io_bookmark, zio, 0, 0); |
b5256303 TC |
526 | } |
527 | } else { | |
528 | zio->io_error = ret; | |
529 | } | |
530 | } | |
531 | ||
b128c09f BB |
532 | /* |
533 | * ========================================================================== | |
534 | * I/O parent/child relationships and pipeline interlocks | |
535 | * ========================================================================== | |
536 | */ | |
d164b209 | 537 | zio_t * |
3dfb57a3 | 538 | zio_walk_parents(zio_t *cio, zio_link_t **zl) |
d164b209 | 539 | { |
d164b209 | 540 | list_t *pl = &cio->io_parent_list; |
b128c09f | 541 | |
3dfb57a3 DB |
542 | *zl = (*zl == NULL) ? list_head(pl) : list_next(pl, *zl); |
543 | if (*zl == NULL) | |
d164b209 BB |
544 | return (NULL); |
545 | ||
3dfb57a3 DB |
546 | ASSERT((*zl)->zl_child == cio); |
547 | return ((*zl)->zl_parent); | |
d164b209 BB |
548 | } |
549 | ||
550 | zio_t * | |
3dfb57a3 | 551 | zio_walk_children(zio_t *pio, zio_link_t **zl) |
d164b209 | 552 | { |
d164b209 BB |
553 | list_t *cl = &pio->io_child_list; |
554 | ||
a8b2e306 TC |
555 | ASSERT(MUTEX_HELD(&pio->io_lock)); |
556 | ||
3dfb57a3 DB |
557 | *zl = (*zl == NULL) ? list_head(cl) : list_next(cl, *zl); |
558 | if (*zl == NULL) | |
d164b209 BB |
559 | return (NULL); |
560 | ||
3dfb57a3 DB |
561 | ASSERT((*zl)->zl_parent == pio); |
562 | return ((*zl)->zl_child); | |
d164b209 BB |
563 | } |
564 | ||
565 | zio_t * | |
566 | zio_unique_parent(zio_t *cio) | |
567 | { | |
3dfb57a3 DB |
568 | zio_link_t *zl = NULL; |
569 | zio_t *pio = zio_walk_parents(cio, &zl); | |
d164b209 | 570 | |
3dfb57a3 | 571 | VERIFY3P(zio_walk_parents(cio, &zl), ==, NULL); |
d164b209 BB |
572 | return (pio); |
573 | } | |
574 | ||
575 | void | |
576 | zio_add_child(zio_t *pio, zio_t *cio) | |
b128c09f | 577 | { |
79c76d5b | 578 | zio_link_t *zl = kmem_cache_alloc(zio_link_cache, KM_SLEEP); |
d164b209 BB |
579 | |
580 | /* | |
581 | * Logical I/Os can have logical, gang, or vdev children. | |
582 | * Gang I/Os can have gang or vdev children. | |
583 | * Vdev I/Os can only have vdev children. | |
584 | * The following ASSERT captures all of these constraints. | |
585 | */ | |
1ce23dca | 586 | ASSERT3S(cio->io_child_type, <=, pio->io_child_type); |
d164b209 BB |
587 | |
588 | zl->zl_parent = pio; | |
589 | zl->zl_child = cio; | |
590 | ||
b128c09f | 591 | mutex_enter(&pio->io_lock); |
a8b2e306 | 592 | mutex_enter(&cio->io_lock); |
d164b209 BB |
593 | |
594 | ASSERT(pio->io_state[ZIO_WAIT_DONE] == 0); | |
595 | ||
1c27024e | 596 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 BB |
597 | pio->io_children[cio->io_child_type][w] += !cio->io_state[w]; |
598 | ||
599 | list_insert_head(&pio->io_child_list, zl); | |
600 | list_insert_head(&cio->io_parent_list, zl); | |
601 | ||
428870ff BB |
602 | pio->io_child_count++; |
603 | cio->io_parent_count++; | |
604 | ||
d164b209 | 605 | mutex_exit(&cio->io_lock); |
a8b2e306 | 606 | mutex_exit(&pio->io_lock); |
b128c09f BB |
607 | } |
608 | ||
34dc7c2f | 609 | static void |
d164b209 | 610 | zio_remove_child(zio_t *pio, zio_t *cio, zio_link_t *zl) |
b128c09f | 611 | { |
d164b209 BB |
612 | ASSERT(zl->zl_parent == pio); |
613 | ASSERT(zl->zl_child == cio); | |
b128c09f BB |
614 | |
615 | mutex_enter(&pio->io_lock); | |
a8b2e306 | 616 | mutex_enter(&cio->io_lock); |
d164b209 BB |
617 | |
618 | list_remove(&pio->io_child_list, zl); | |
619 | list_remove(&cio->io_parent_list, zl); | |
620 | ||
428870ff BB |
621 | pio->io_child_count--; |
622 | cio->io_parent_count--; | |
623 | ||
d164b209 | 624 | mutex_exit(&cio->io_lock); |
a8b2e306 | 625 | mutex_exit(&pio->io_lock); |
d164b209 | 626 | kmem_cache_free(zio_link_cache, zl); |
b128c09f BB |
627 | } |
628 | ||
629 | static boolean_t | |
ddc751d5 | 630 | zio_wait_for_children(zio_t *zio, uint8_t childbits, enum zio_wait_type wait) |
34dc7c2f | 631 | { |
b128c09f BB |
632 | boolean_t waiting = B_FALSE; |
633 | ||
634 | mutex_enter(&zio->io_lock); | |
635 | ASSERT(zio->io_stall == NULL); | |
ddc751d5 GW |
636 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) { |
637 | if (!(ZIO_CHILD_BIT_IS_SET(childbits, c))) | |
638 | continue; | |
639 | ||
640 | uint64_t *countp = &zio->io_children[c][wait]; | |
641 | if (*countp != 0) { | |
642 | zio->io_stage >>= 1; | |
643 | ASSERT3U(zio->io_stage, !=, ZIO_STAGE_OPEN); | |
644 | zio->io_stall = countp; | |
645 | waiting = B_TRUE; | |
646 | break; | |
647 | } | |
b128c09f BB |
648 | } |
649 | mutex_exit(&zio->io_lock); | |
b128c09f BB |
650 | return (waiting); |
651 | } | |
34dc7c2f | 652 | |
bf701a83 BB |
653 | __attribute__((always_inline)) |
654 | static inline void | |
62840030 MA |
655 | zio_notify_parent(zio_t *pio, zio_t *zio, enum zio_wait_type wait, |
656 | zio_t **next_to_executep) | |
b128c09f BB |
657 | { |
658 | uint64_t *countp = &pio->io_children[zio->io_child_type][wait]; | |
659 | int *errorp = &pio->io_child_error[zio->io_child_type]; | |
34dc7c2f | 660 | |
b128c09f BB |
661 | mutex_enter(&pio->io_lock); |
662 | if (zio->io_error && !(zio->io_flags & ZIO_FLAG_DONT_PROPAGATE)) | |
663 | *errorp = zio_worst_error(*errorp, zio->io_error); | |
664 | pio->io_reexecute |= zio->io_reexecute; | |
665 | ASSERT3U(*countp, >, 0); | |
e8b96c60 MA |
666 | |
667 | (*countp)--; | |
668 | ||
669 | if (*countp == 0 && pio->io_stall == countp) { | |
3dfb57a3 DB |
670 | zio_taskq_type_t type = |
671 | pio->io_stage < ZIO_STAGE_VDEV_IO_START ? ZIO_TASKQ_ISSUE : | |
672 | ZIO_TASKQ_INTERRUPT; | |
b128c09f BB |
673 | pio->io_stall = NULL; |
674 | mutex_exit(&pio->io_lock); | |
62840030 | 675 | |
3dfb57a3 | 676 | /* |
62840030 MA |
677 | * If we can tell the caller to execute this parent next, do |
678 | * so. Otherwise dispatch the parent zio as its own task. | |
679 | * | |
680 | * Having the caller execute the parent when possible reduces | |
681 | * locking on the zio taskq's, reduces context switch | |
682 | * overhead, and has no recursion penalty. Note that one | |
683 | * read from disk typically causes at least 3 zio's: a | |
684 | * zio_null(), the logical zio_read(), and then a physical | |
685 | * zio. When the physical ZIO completes, we are able to call | |
686 | * zio_done() on all 3 of these zio's from one invocation of | |
687 | * zio_execute() by returning the parent back to | |
688 | * zio_execute(). Since the parent isn't executed until this | |
689 | * thread returns back to zio_execute(), the caller should do | |
690 | * so promptly. | |
691 | * | |
692 | * In other cases, dispatching the parent prevents | |
693 | * overflowing the stack when we have deeply nested | |
694 | * parent-child relationships, as we do with the "mega zio" | |
695 | * of writes for spa_sync(), and the chain of ZIL blocks. | |
3dfb57a3 | 696 | */ |
62840030 MA |
697 | if (next_to_executep != NULL && *next_to_executep == NULL) { |
698 | *next_to_executep = pio; | |
699 | } else { | |
700 | zio_taskq_dispatch(pio, type, B_FALSE); | |
701 | } | |
b128c09f BB |
702 | } else { |
703 | mutex_exit(&pio->io_lock); | |
34dc7c2f BB |
704 | } |
705 | } | |
706 | ||
b128c09f BB |
707 | static void |
708 | zio_inherit_child_errors(zio_t *zio, enum zio_child c) | |
709 | { | |
710 | if (zio->io_child_error[c] != 0 && zio->io_error == 0) | |
711 | zio->io_error = zio->io_child_error[c]; | |
712 | } | |
713 | ||
3dfb57a3 | 714 | int |
64fc7762 | 715 | zio_bookmark_compare(const void *x1, const void *x2) |
3dfb57a3 DB |
716 | { |
717 | const zio_t *z1 = x1; | |
718 | const zio_t *z2 = x2; | |
3dfb57a3 | 719 | |
64fc7762 MA |
720 | if (z1->io_bookmark.zb_objset < z2->io_bookmark.zb_objset) |
721 | return (-1); | |
722 | if (z1->io_bookmark.zb_objset > z2->io_bookmark.zb_objset) | |
723 | return (1); | |
3dfb57a3 | 724 | |
64fc7762 MA |
725 | if (z1->io_bookmark.zb_object < z2->io_bookmark.zb_object) |
726 | return (-1); | |
727 | if (z1->io_bookmark.zb_object > z2->io_bookmark.zb_object) | |
728 | return (1); | |
3dfb57a3 | 729 | |
64fc7762 MA |
730 | if (z1->io_bookmark.zb_level < z2->io_bookmark.zb_level) |
731 | return (-1); | |
732 | if (z1->io_bookmark.zb_level > z2->io_bookmark.zb_level) | |
733 | return (1); | |
734 | ||
735 | if (z1->io_bookmark.zb_blkid < z2->io_bookmark.zb_blkid) | |
736 | return (-1); | |
737 | if (z1->io_bookmark.zb_blkid > z2->io_bookmark.zb_blkid) | |
738 | return (1); | |
739 | ||
740 | if (z1 < z2) | |
741 | return (-1); | |
742 | if (z1 > z2) | |
743 | return (1); | |
744 | ||
745 | return (0); | |
3dfb57a3 DB |
746 | } |
747 | ||
34dc7c2f BB |
748 | /* |
749 | * ========================================================================== | |
b128c09f | 750 | * Create the various types of I/O (read, write, free, etc) |
34dc7c2f BB |
751 | * ========================================================================== |
752 | */ | |
753 | static zio_t * | |
428870ff | 754 | zio_create(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
a6255b7f | 755 | abd_t *data, uint64_t lsize, uint64_t psize, zio_done_func_t *done, |
2aa34383 DK |
756 | void *private, zio_type_t type, zio_priority_t priority, |
757 | enum zio_flag flags, vdev_t *vd, uint64_t offset, | |
758 | const zbookmark_phys_t *zb, enum zio_stage stage, | |
759 | enum zio_stage pipeline) | |
34dc7c2f BB |
760 | { |
761 | zio_t *zio; | |
762 | ||
2aa34383 DK |
763 | ASSERT3U(psize, <=, SPA_MAXBLOCKSIZE); |
764 | ASSERT(P2PHASE(psize, SPA_MINBLOCKSIZE) == 0); | |
b128c09f BB |
765 | ASSERT(P2PHASE(offset, SPA_MINBLOCKSIZE) == 0); |
766 | ||
767 | ASSERT(!vd || spa_config_held(spa, SCL_STATE_ALL, RW_READER)); | |
768 | ASSERT(!bp || !(flags & ZIO_FLAG_CONFIG_WRITER)); | |
769 | ASSERT(vd || stage == ZIO_STAGE_OPEN); | |
34dc7c2f | 770 | |
b5256303 | 771 | IMPLY(lsize != psize, (flags & ZIO_FLAG_RAW_COMPRESS) != 0); |
2aa34383 | 772 | |
79c76d5b | 773 | zio = kmem_cache_alloc(zio_cache, KM_SLEEP); |
3941503c BB |
774 | bzero(zio, sizeof (zio_t)); |
775 | ||
448d7aaa | 776 | mutex_init(&zio->io_lock, NULL, MUTEX_NOLOCKDEP, NULL); |
3941503c BB |
777 | cv_init(&zio->io_cv, NULL, CV_DEFAULT, NULL); |
778 | ||
779 | list_create(&zio->io_parent_list, sizeof (zio_link_t), | |
780 | offsetof(zio_link_t, zl_parent_node)); | |
781 | list_create(&zio->io_child_list, sizeof (zio_link_t), | |
782 | offsetof(zio_link_t, zl_child_node)); | |
4e21fd06 | 783 | metaslab_trace_init(&zio->io_alloc_list); |
d164b209 | 784 | |
b128c09f BB |
785 | if (vd != NULL) |
786 | zio->io_child_type = ZIO_CHILD_VDEV; | |
787 | else if (flags & ZIO_FLAG_GANG_CHILD) | |
788 | zio->io_child_type = ZIO_CHILD_GANG; | |
428870ff BB |
789 | else if (flags & ZIO_FLAG_DDT_CHILD) |
790 | zio->io_child_type = ZIO_CHILD_DDT; | |
b128c09f BB |
791 | else |
792 | zio->io_child_type = ZIO_CHILD_LOGICAL; | |
793 | ||
34dc7c2f | 794 | if (bp != NULL) { |
428870ff | 795 | zio->io_bp = (blkptr_t *)bp; |
34dc7c2f BB |
796 | zio->io_bp_copy = *bp; |
797 | zio->io_bp_orig = *bp; | |
428870ff BB |
798 | if (type != ZIO_TYPE_WRITE || |
799 | zio->io_child_type == ZIO_CHILD_DDT) | |
b128c09f | 800 | zio->io_bp = &zio->io_bp_copy; /* so caller can free */ |
9babb374 | 801 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) |
b128c09f | 802 | zio->io_logical = zio; |
9babb374 BB |
803 | if (zio->io_child_type > ZIO_CHILD_GANG && BP_IS_GANG(bp)) |
804 | pipeline |= ZIO_GANG_STAGES; | |
34dc7c2f | 805 | } |
b128c09f BB |
806 | |
807 | zio->io_spa = spa; | |
808 | zio->io_txg = txg; | |
34dc7c2f BB |
809 | zio->io_done = done; |
810 | zio->io_private = private; | |
811 | zio->io_type = type; | |
812 | zio->io_priority = priority; | |
b128c09f BB |
813 | zio->io_vd = vd; |
814 | zio->io_offset = offset; | |
a6255b7f | 815 | zio->io_orig_abd = zio->io_abd = data; |
2aa34383 DK |
816 | zio->io_orig_size = zio->io_size = psize; |
817 | zio->io_lsize = lsize; | |
b128c09f BB |
818 | zio->io_orig_flags = zio->io_flags = flags; |
819 | zio->io_orig_stage = zio->io_stage = stage; | |
820 | zio->io_orig_pipeline = zio->io_pipeline = pipeline; | |
3dfb57a3 | 821 | zio->io_pipeline_trace = ZIO_STAGE_OPEN; |
34dc7c2f | 822 | |
d164b209 BB |
823 | zio->io_state[ZIO_WAIT_READY] = (stage >= ZIO_STAGE_READY); |
824 | zio->io_state[ZIO_WAIT_DONE] = (stage >= ZIO_STAGE_DONE); | |
825 | ||
b128c09f BB |
826 | if (zb != NULL) |
827 | zio->io_bookmark = *zb; | |
828 | ||
829 | if (pio != NULL) { | |
cc99f275 DB |
830 | if (zio->io_metaslab_class == NULL) |
831 | zio->io_metaslab_class = pio->io_metaslab_class; | |
b128c09f | 832 | if (zio->io_logical == NULL) |
34dc7c2f | 833 | zio->io_logical = pio->io_logical; |
9babb374 BB |
834 | if (zio->io_child_type == ZIO_CHILD_GANG) |
835 | zio->io_gang_leader = pio->io_gang_leader; | |
b128c09f | 836 | zio_add_child(pio, zio); |
34dc7c2f BB |
837 | } |
838 | ||
a38718a6 GA |
839 | taskq_init_ent(&zio->io_tqent); |
840 | ||
34dc7c2f BB |
841 | return (zio); |
842 | } | |
843 | ||
844 | static void | |
b128c09f | 845 | zio_destroy(zio_t *zio) |
34dc7c2f | 846 | { |
4e21fd06 | 847 | metaslab_trace_fini(&zio->io_alloc_list); |
3941503c BB |
848 | list_destroy(&zio->io_parent_list); |
849 | list_destroy(&zio->io_child_list); | |
850 | mutex_destroy(&zio->io_lock); | |
851 | cv_destroy(&zio->io_cv); | |
b128c09f | 852 | kmem_cache_free(zio_cache, zio); |
34dc7c2f BB |
853 | } |
854 | ||
855 | zio_t * | |
d164b209 | 856 | zio_null(zio_t *pio, spa_t *spa, vdev_t *vd, zio_done_func_t *done, |
428870ff | 857 | void *private, enum zio_flag flags) |
34dc7c2f BB |
858 | { |
859 | zio_t *zio; | |
860 | ||
2aa34383 | 861 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
d164b209 | 862 | ZIO_TYPE_NULL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
b128c09f | 863 | ZIO_STAGE_OPEN, ZIO_INTERLOCK_PIPELINE); |
34dc7c2f BB |
864 | |
865 | return (zio); | |
866 | } | |
867 | ||
868 | zio_t * | |
428870ff | 869 | zio_root(spa_t *spa, zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f | 870 | { |
d164b209 | 871 | return (zio_null(NULL, spa, NULL, done, private, flags)); |
34dc7c2f BB |
872 | } |
873 | ||
63e3a861 MA |
874 | void |
875 | zfs_blkptr_verify(spa_t *spa, const blkptr_t *bp) | |
876 | { | |
63e3a861 MA |
877 | if (!DMU_OT_IS_VALID(BP_GET_TYPE(bp))) { |
878 | zfs_panic_recover("blkptr at %p has invalid TYPE %llu", | |
879 | bp, (longlong_t)BP_GET_TYPE(bp)); | |
880 | } | |
881 | if (BP_GET_CHECKSUM(bp) >= ZIO_CHECKSUM_FUNCTIONS || | |
882 | BP_GET_CHECKSUM(bp) <= ZIO_CHECKSUM_ON) { | |
883 | zfs_panic_recover("blkptr at %p has invalid CHECKSUM %llu", | |
884 | bp, (longlong_t)BP_GET_CHECKSUM(bp)); | |
885 | } | |
886 | if (BP_GET_COMPRESS(bp) >= ZIO_COMPRESS_FUNCTIONS || | |
887 | BP_GET_COMPRESS(bp) <= ZIO_COMPRESS_ON) { | |
888 | zfs_panic_recover("blkptr at %p has invalid COMPRESS %llu", | |
889 | bp, (longlong_t)BP_GET_COMPRESS(bp)); | |
890 | } | |
891 | if (BP_GET_LSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
892 | zfs_panic_recover("blkptr at %p has invalid LSIZE %llu", | |
893 | bp, (longlong_t)BP_GET_LSIZE(bp)); | |
894 | } | |
895 | if (BP_GET_PSIZE(bp) > SPA_MAXBLOCKSIZE) { | |
896 | zfs_panic_recover("blkptr at %p has invalid PSIZE %llu", | |
897 | bp, (longlong_t)BP_GET_PSIZE(bp)); | |
898 | } | |
899 | ||
900 | if (BP_IS_EMBEDDED(bp)) { | |
901 | if (BPE_GET_ETYPE(bp) > NUM_BP_EMBEDDED_TYPES) { | |
902 | zfs_panic_recover("blkptr at %p has invalid ETYPE %llu", | |
903 | bp, (longlong_t)BPE_GET_ETYPE(bp)); | |
904 | } | |
905 | } | |
906 | ||
6cb8e530 PZ |
907 | /* |
908 | * Do not verify individual DVAs if the config is not trusted. This | |
909 | * will be done once the zio is executed in vdev_mirror_map_alloc. | |
910 | */ | |
911 | if (!spa->spa_trust_config) | |
912 | return; | |
913 | ||
63e3a861 MA |
914 | /* |
915 | * Pool-specific checks. | |
916 | * | |
917 | * Note: it would be nice to verify that the blk_birth and | |
918 | * BP_PHYSICAL_BIRTH() are not too large. However, spa_freeze() | |
919 | * allows the birth time of log blocks (and dmu_sync()-ed blocks | |
920 | * that are in the log) to be arbitrarily large. | |
921 | */ | |
1c27024e | 922 | for (int i = 0; i < BP_GET_NDVAS(bp); i++) { |
63e3a861 | 923 | uint64_t vdevid = DVA_GET_VDEV(&bp->blk_dva[i]); |
1c27024e | 924 | |
63e3a861 MA |
925 | if (vdevid >= spa->spa_root_vdev->vdev_children) { |
926 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
927 | "VDEV %llu", | |
928 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 929 | continue; |
63e3a861 | 930 | } |
1c27024e | 931 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; |
63e3a861 MA |
932 | if (vd == NULL) { |
933 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
934 | "VDEV %llu", | |
935 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 936 | continue; |
63e3a861 MA |
937 | } |
938 | if (vd->vdev_ops == &vdev_hole_ops) { | |
939 | zfs_panic_recover("blkptr at %p DVA %u has hole " | |
940 | "VDEV %llu", | |
941 | bp, i, (longlong_t)vdevid); | |
ee3a23b8 | 942 | continue; |
63e3a861 MA |
943 | } |
944 | if (vd->vdev_ops == &vdev_missing_ops) { | |
945 | /* | |
946 | * "missing" vdevs are valid during import, but we | |
947 | * don't have their detailed info (e.g. asize), so | |
948 | * we can't perform any more checks on them. | |
949 | */ | |
950 | continue; | |
951 | } | |
1c27024e DB |
952 | uint64_t offset = DVA_GET_OFFSET(&bp->blk_dva[i]); |
953 | uint64_t asize = DVA_GET_ASIZE(&bp->blk_dva[i]); | |
63e3a861 MA |
954 | if (BP_IS_GANG(bp)) |
955 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
956 | if (offset + asize > vd->vdev_asize) { | |
957 | zfs_panic_recover("blkptr at %p DVA %u has invalid " | |
958 | "OFFSET %llu", | |
959 | bp, i, (longlong_t)offset); | |
960 | } | |
961 | } | |
962 | } | |
963 | ||
6cb8e530 PZ |
964 | boolean_t |
965 | zfs_dva_valid(spa_t *spa, const dva_t *dva, const blkptr_t *bp) | |
966 | { | |
967 | uint64_t vdevid = DVA_GET_VDEV(dva); | |
968 | ||
969 | if (vdevid >= spa->spa_root_vdev->vdev_children) | |
970 | return (B_FALSE); | |
971 | ||
972 | vdev_t *vd = spa->spa_root_vdev->vdev_child[vdevid]; | |
973 | if (vd == NULL) | |
974 | return (B_FALSE); | |
975 | ||
976 | if (vd->vdev_ops == &vdev_hole_ops) | |
977 | return (B_FALSE); | |
978 | ||
979 | if (vd->vdev_ops == &vdev_missing_ops) { | |
980 | return (B_FALSE); | |
981 | } | |
982 | ||
983 | uint64_t offset = DVA_GET_OFFSET(dva); | |
984 | uint64_t asize = DVA_GET_ASIZE(dva); | |
985 | ||
986 | if (BP_IS_GANG(bp)) | |
987 | asize = vdev_psize_to_asize(vd, SPA_GANGBLOCKSIZE); | |
988 | if (offset + asize > vd->vdev_asize) | |
989 | return (B_FALSE); | |
990 | ||
991 | return (B_TRUE); | |
992 | } | |
993 | ||
34dc7c2f | 994 | zio_t * |
b128c09f | 995 | zio_read(zio_t *pio, spa_t *spa, const blkptr_t *bp, |
a6255b7f | 996 | abd_t *data, uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 997 | zio_priority_t priority, enum zio_flag flags, const zbookmark_phys_t *zb) |
34dc7c2f BB |
998 | { |
999 | zio_t *zio; | |
1000 | ||
63e3a861 MA |
1001 | zfs_blkptr_verify(spa, bp); |
1002 | ||
428870ff | 1003 | zio = zio_create(pio, spa, BP_PHYSICAL_BIRTH(bp), bp, |
2aa34383 | 1004 | data, size, size, done, private, |
b128c09f | 1005 | ZIO_TYPE_READ, priority, flags, NULL, 0, zb, |
428870ff BB |
1006 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1007 | ZIO_DDT_CHILD_READ_PIPELINE : ZIO_READ_PIPELINE); | |
34dc7c2f | 1008 | |
b128c09f BB |
1009 | return (zio); |
1010 | } | |
34dc7c2f | 1011 | |
34dc7c2f | 1012 | zio_t * |
b128c09f | 1013 | zio_write(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, |
a6255b7f | 1014 | abd_t *data, uint64_t lsize, uint64_t psize, const zio_prop_t *zp, |
bc77ba73 PD |
1015 | zio_done_func_t *ready, zio_done_func_t *children_ready, |
1016 | zio_done_func_t *physdone, zio_done_func_t *done, | |
1017 | void *private, zio_priority_t priority, enum zio_flag flags, | |
1018 | const zbookmark_phys_t *zb) | |
34dc7c2f BB |
1019 | { |
1020 | zio_t *zio; | |
1021 | ||
b128c09f BB |
1022 | ASSERT(zp->zp_checksum >= ZIO_CHECKSUM_OFF && |
1023 | zp->zp_checksum < ZIO_CHECKSUM_FUNCTIONS && | |
1024 | zp->zp_compress >= ZIO_COMPRESS_OFF && | |
1025 | zp->zp_compress < ZIO_COMPRESS_FUNCTIONS && | |
9ae529ec | 1026 | DMU_OT_IS_VALID(zp->zp_type) && |
b128c09f | 1027 | zp->zp_level < 32 && |
428870ff | 1028 | zp->zp_copies > 0 && |
03c6040b | 1029 | zp->zp_copies <= spa_max_replication(spa)); |
34dc7c2f | 1030 | |
2aa34383 | 1031 | zio = zio_create(pio, spa, txg, bp, data, lsize, psize, done, private, |
b128c09f | 1032 | ZIO_TYPE_WRITE, priority, flags, NULL, 0, zb, |
428870ff BB |
1033 | ZIO_STAGE_OPEN, (flags & ZIO_FLAG_DDT_CHILD) ? |
1034 | ZIO_DDT_CHILD_WRITE_PIPELINE : ZIO_WRITE_PIPELINE); | |
34dc7c2f BB |
1035 | |
1036 | zio->io_ready = ready; | |
bc77ba73 | 1037 | zio->io_children_ready = children_ready; |
e8b96c60 | 1038 | zio->io_physdone = physdone; |
b128c09f | 1039 | zio->io_prop = *zp; |
34dc7c2f | 1040 | |
9b67f605 MA |
1041 | /* |
1042 | * Data can be NULL if we are going to call zio_write_override() to | |
1043 | * provide the already-allocated BP. But we may need the data to | |
1044 | * verify a dedup hit (if requested). In this case, don't try to | |
b5256303 TC |
1045 | * dedup (just take the already-allocated BP verbatim). Encrypted |
1046 | * dedup blocks need data as well so we also disable dedup in this | |
1047 | * case. | |
9b67f605 | 1048 | */ |
b5256303 TC |
1049 | if (data == NULL && |
1050 | (zio->io_prop.zp_dedup_verify || zio->io_prop.zp_encrypt)) { | |
9b67f605 MA |
1051 | zio->io_prop.zp_dedup = zio->io_prop.zp_dedup_verify = B_FALSE; |
1052 | } | |
1053 | ||
34dc7c2f BB |
1054 | return (zio); |
1055 | } | |
1056 | ||
1057 | zio_t * | |
a6255b7f | 1058 | zio_rewrite(zio_t *pio, spa_t *spa, uint64_t txg, blkptr_t *bp, abd_t *data, |
e8b96c60 | 1059 | uint64_t size, zio_done_func_t *done, void *private, |
5dbd68a3 | 1060 | zio_priority_t priority, enum zio_flag flags, zbookmark_phys_t *zb) |
34dc7c2f BB |
1061 | { |
1062 | zio_t *zio; | |
1063 | ||
2aa34383 | 1064 | zio = zio_create(pio, spa, txg, bp, data, size, size, done, private, |
3dfb57a3 | 1065 | ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_IO_REWRITE, NULL, 0, zb, |
b128c09f | 1066 | ZIO_STAGE_OPEN, ZIO_REWRITE_PIPELINE); |
34dc7c2f BB |
1067 | |
1068 | return (zio); | |
1069 | } | |
1070 | ||
428870ff | 1071 | void |
03c6040b | 1072 | zio_write_override(zio_t *zio, blkptr_t *bp, int copies, boolean_t nopwrite) |
428870ff BB |
1073 | { |
1074 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
1075 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1076 | ASSERT(zio->io_stage == ZIO_STAGE_OPEN); | |
1077 | ASSERT(zio->io_txg == spa_syncing_txg(zio->io_spa)); | |
1078 | ||
03c6040b GW |
1079 | /* |
1080 | * We must reset the io_prop to match the values that existed | |
1081 | * when the bp was first written by dmu_sync() keeping in mind | |
1082 | * that nopwrite and dedup are mutually exclusive. | |
1083 | */ | |
1084 | zio->io_prop.zp_dedup = nopwrite ? B_FALSE : zio->io_prop.zp_dedup; | |
1085 | zio->io_prop.zp_nopwrite = nopwrite; | |
428870ff BB |
1086 | zio->io_prop.zp_copies = copies; |
1087 | zio->io_bp_override = bp; | |
1088 | } | |
1089 | ||
1090 | void | |
1091 | zio_free(spa_t *spa, uint64_t txg, const blkptr_t *bp) | |
1092 | { | |
9b67f605 | 1093 | |
a1d477c2 MA |
1094 | zfs_blkptr_verify(spa, bp); |
1095 | ||
9b67f605 MA |
1096 | /* |
1097 | * The check for EMBEDDED is a performance optimization. We | |
1098 | * process the free here (by ignoring it) rather than | |
1099 | * putting it on the list and then processing it in zio_free_sync(). | |
1100 | */ | |
1101 | if (BP_IS_EMBEDDED(bp)) | |
1102 | return; | |
13fe0198 | 1103 | metaslab_check_free(spa, bp); |
2883cad5 MA |
1104 | |
1105 | /* | |
1106 | * Frees that are for the currently-syncing txg, are not going to be | |
1107 | * deferred, and which will not need to do a read (i.e. not GANG or | |
1108 | * DEDUP), can be processed immediately. Otherwise, put them on the | |
1109 | * in-memory list for later processing. | |
1110 | */ | |
1111 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp) || | |
1112 | txg != spa->spa_syncing_txg || | |
1113 | spa_sync_pass(spa) >= zfs_sync_pass_deferred_free) { | |
1114 | bplist_append(&spa->spa_free_bplist[txg & TXG_MASK], bp); | |
1115 | } else { | |
1116 | VERIFY0(zio_wait(zio_free_sync(NULL, spa, txg, bp, 0))); | |
1117 | } | |
428870ff BB |
1118 | } |
1119 | ||
34dc7c2f | 1120 | zio_t * |
428870ff BB |
1121 | zio_free_sync(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1122 | enum zio_flag flags) | |
34dc7c2f BB |
1123 | { |
1124 | zio_t *zio; | |
2883cad5 | 1125 | enum zio_stage stage = ZIO_FREE_PIPELINE; |
34dc7c2f | 1126 | |
428870ff BB |
1127 | ASSERT(!BP_IS_HOLE(bp)); |
1128 | ASSERT(spa_syncing_txg(spa) == txg); | |
55d85d5a | 1129 | ASSERT(spa_sync_pass(spa) < zfs_sync_pass_deferred_free); |
34dc7c2f | 1130 | |
9b67f605 MA |
1131 | if (BP_IS_EMBEDDED(bp)) |
1132 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1133 | ||
13fe0198 | 1134 | metaslab_check_free(spa, bp); |
8c841793 | 1135 | arc_freed(spa, bp); |
d4a72f23 | 1136 | dsl_scan_freed(spa, bp); |
13fe0198 | 1137 | |
2883cad5 MA |
1138 | /* |
1139 | * GANG and DEDUP blocks can induce a read (for the gang block header, | |
1140 | * or the DDT), so issue them asynchronously so that this thread is | |
1141 | * not tied up. | |
1142 | */ | |
1143 | if (BP_IS_GANG(bp) || BP_GET_DEDUP(bp)) | |
1144 | stage |= ZIO_STAGE_ISSUE_ASYNC; | |
1145 | ||
b128c09f | 1146 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1147 | BP_GET_PSIZE(bp), NULL, NULL, ZIO_TYPE_FREE, ZIO_PRIORITY_NOW, |
1148 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, stage); | |
2883cad5 | 1149 | |
34dc7c2f BB |
1150 | return (zio); |
1151 | } | |
1152 | ||
1153 | zio_t * | |
428870ff BB |
1154 | zio_claim(zio_t *pio, spa_t *spa, uint64_t txg, const blkptr_t *bp, |
1155 | zio_done_func_t *done, void *private, enum zio_flag flags) | |
34dc7c2f BB |
1156 | { |
1157 | zio_t *zio; | |
1158 | ||
a1d477c2 | 1159 | zfs_blkptr_verify(spa, bp); |
9b67f605 MA |
1160 | |
1161 | if (BP_IS_EMBEDDED(bp)) | |
1162 | return (zio_null(pio, spa, NULL, NULL, NULL, 0)); | |
1163 | ||
34dc7c2f BB |
1164 | /* |
1165 | * A claim is an allocation of a specific block. Claims are needed | |
1166 | * to support immediate writes in the intent log. The issue is that | |
1167 | * immediate writes contain committed data, but in a txg that was | |
1168 | * *not* committed. Upon opening the pool after an unclean shutdown, | |
1169 | * the intent log claims all blocks that contain immediate write data | |
1170 | * so that the SPA knows they're in use. | |
1171 | * | |
1172 | * All claims *must* be resolved in the first txg -- before the SPA | |
1173 | * starts allocating blocks -- so that nothing is allocated twice. | |
428870ff | 1174 | * If txg == 0 we just verify that the block is claimable. |
34dc7c2f | 1175 | */ |
d2734cce SD |
1176 | ASSERT3U(spa->spa_uberblock.ub_rootbp.blk_birth, <, |
1177 | spa_min_claim_txg(spa)); | |
1178 | ASSERT(txg == spa_min_claim_txg(spa) || txg == 0); | |
428870ff | 1179 | ASSERT(!BP_GET_DEDUP(bp) || !spa_writeable(spa)); /* zdb(1M) */ |
34dc7c2f | 1180 | |
b128c09f | 1181 | zio = zio_create(pio, spa, txg, bp, NULL, BP_GET_PSIZE(bp), |
2aa34383 DK |
1182 | BP_GET_PSIZE(bp), done, private, ZIO_TYPE_CLAIM, ZIO_PRIORITY_NOW, |
1183 | flags, NULL, 0, NULL, ZIO_STAGE_OPEN, ZIO_CLAIM_PIPELINE); | |
3dfb57a3 | 1184 | ASSERT0(zio->io_queued_timestamp); |
34dc7c2f BB |
1185 | |
1186 | return (zio); | |
1187 | } | |
1188 | ||
1189 | zio_t * | |
1190 | zio_ioctl(zio_t *pio, spa_t *spa, vdev_t *vd, int cmd, | |
e8b96c60 | 1191 | zio_done_func_t *done, void *private, enum zio_flag flags) |
34dc7c2f BB |
1192 | { |
1193 | zio_t *zio; | |
1194 | int c; | |
1195 | ||
1196 | if (vd->vdev_children == 0) { | |
2aa34383 | 1197 | zio = zio_create(pio, spa, 0, NULL, NULL, 0, 0, done, private, |
e8b96c60 | 1198 | ZIO_TYPE_IOCTL, ZIO_PRIORITY_NOW, flags, vd, 0, NULL, |
34dc7c2f BB |
1199 | ZIO_STAGE_OPEN, ZIO_IOCTL_PIPELINE); |
1200 | ||
34dc7c2f BB |
1201 | zio->io_cmd = cmd; |
1202 | } else { | |
d164b209 | 1203 | zio = zio_null(pio, spa, NULL, NULL, NULL, flags); |
34dc7c2f BB |
1204 | |
1205 | for (c = 0; c < vd->vdev_children; c++) | |
1206 | zio_nowait(zio_ioctl(zio, spa, vd->vdev_child[c], cmd, | |
e8b96c60 | 1207 | done, private, flags)); |
34dc7c2f BB |
1208 | } |
1209 | ||
1210 | return (zio); | |
1211 | } | |
1212 | ||
34dc7c2f BB |
1213 | zio_t * |
1214 | zio_read_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1215 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1216 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f BB |
1217 | { |
1218 | zio_t *zio; | |
34dc7c2f | 1219 | |
b128c09f BB |
1220 | ASSERT(vd->vdev_children == 0); |
1221 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1222 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1223 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1224 | |
2aa34383 DK |
1225 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1226 | private, ZIO_TYPE_READ, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1227 | offset, NULL, ZIO_STAGE_OPEN, ZIO_READ_PHYS_PIPELINE); | |
34dc7c2f | 1228 | |
b128c09f | 1229 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f BB |
1230 | |
1231 | return (zio); | |
1232 | } | |
1233 | ||
1234 | zio_t * | |
1235 | zio_write_phys(zio_t *pio, vdev_t *vd, uint64_t offset, uint64_t size, | |
a6255b7f | 1236 | abd_t *data, int checksum, zio_done_func_t *done, void *private, |
e8b96c60 | 1237 | zio_priority_t priority, enum zio_flag flags, boolean_t labels) |
34dc7c2f | 1238 | { |
34dc7c2f | 1239 | zio_t *zio; |
34dc7c2f | 1240 | |
b128c09f BB |
1241 | ASSERT(vd->vdev_children == 0); |
1242 | ASSERT(!labels || offset + size <= VDEV_LABEL_START_SIZE || | |
1243 | offset >= vd->vdev_psize - VDEV_LABEL_END_SIZE); | |
1244 | ASSERT3U(offset + size, <=, vd->vdev_psize); | |
34dc7c2f | 1245 | |
2aa34383 DK |
1246 | zio = zio_create(pio, vd->vdev_spa, 0, NULL, data, size, size, done, |
1247 | private, ZIO_TYPE_WRITE, priority, flags | ZIO_FLAG_PHYSICAL, vd, | |
1248 | offset, NULL, ZIO_STAGE_OPEN, ZIO_WRITE_PHYS_PIPELINE); | |
34dc7c2f | 1249 | |
b128c09f | 1250 | zio->io_prop.zp_checksum = checksum; |
34dc7c2f | 1251 | |
3c67d83a | 1252 | if (zio_checksum_table[checksum].ci_flags & ZCHECKSUM_FLAG_EMBEDDED) { |
34dc7c2f | 1253 | /* |
428870ff | 1254 | * zec checksums are necessarily destructive -- they modify |
b128c09f | 1255 | * the end of the write buffer to hold the verifier/checksum. |
34dc7c2f | 1256 | * Therefore, we must make a local copy in case the data is |
b128c09f | 1257 | * being written to multiple places in parallel. |
34dc7c2f | 1258 | */ |
a6255b7f DQ |
1259 | abd_t *wbuf = abd_alloc_sametype(data, size); |
1260 | abd_copy(wbuf, data, size); | |
1261 | ||
b128c09f | 1262 | zio_push_transform(zio, wbuf, size, size, NULL); |
34dc7c2f BB |
1263 | } |
1264 | ||
1265 | return (zio); | |
1266 | } | |
1267 | ||
1268 | /* | |
b128c09f | 1269 | * Create a child I/O to do some work for us. |
34dc7c2f BB |
1270 | */ |
1271 | zio_t * | |
b128c09f | 1272 | zio_vdev_child_io(zio_t *pio, blkptr_t *bp, vdev_t *vd, uint64_t offset, |
4ea3f864 GM |
1273 | abd_t *data, uint64_t size, int type, zio_priority_t priority, |
1274 | enum zio_flag flags, zio_done_func_t *done, void *private) | |
34dc7c2f | 1275 | { |
428870ff | 1276 | enum zio_stage pipeline = ZIO_VDEV_CHILD_PIPELINE; |
b128c09f BB |
1277 | zio_t *zio; |
1278 | ||
a1d477c2 MA |
1279 | /* |
1280 | * vdev child I/Os do not propagate their error to the parent. | |
1281 | * Therefore, for correct operation the caller *must* check for | |
1282 | * and handle the error in the child i/o's done callback. | |
1283 | * The only exceptions are i/os that we don't care about | |
1284 | * (OPTIONAL or REPAIR). | |
1285 | */ | |
1286 | ASSERT((flags & ZIO_FLAG_OPTIONAL) || (flags & ZIO_FLAG_IO_REPAIR) || | |
1287 | done != NULL); | |
1288 | ||
34dc7c2f BB |
1289 | if (type == ZIO_TYPE_READ && bp != NULL) { |
1290 | /* | |
1291 | * If we have the bp, then the child should perform the | |
1292 | * checksum and the parent need not. This pushes error | |
1293 | * detection as close to the leaves as possible and | |
1294 | * eliminates redundant checksums in the interior nodes. | |
1295 | */ | |
428870ff BB |
1296 | pipeline |= ZIO_STAGE_CHECKSUM_VERIFY; |
1297 | pio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; | |
34dc7c2f BB |
1298 | } |
1299 | ||
a1d477c2 MA |
1300 | if (vd->vdev_ops->vdev_op_leaf) { |
1301 | ASSERT0(vd->vdev_children); | |
b128c09f | 1302 | offset += VDEV_LABEL_START_SIZE; |
a1d477c2 | 1303 | } |
b128c09f | 1304 | |
a1d477c2 | 1305 | flags |= ZIO_VDEV_CHILD_FLAGS(pio); |
428870ff BB |
1306 | |
1307 | /* | |
1308 | * If we've decided to do a repair, the write is not speculative -- | |
1309 | * even if the original read was. | |
1310 | */ | |
1311 | if (flags & ZIO_FLAG_IO_REPAIR) | |
1312 | flags &= ~ZIO_FLAG_SPECULATIVE; | |
1313 | ||
3dfb57a3 DB |
1314 | /* |
1315 | * If we're creating a child I/O that is not associated with a | |
1316 | * top-level vdev, then the child zio is not an allocating I/O. | |
1317 | * If this is a retried I/O then we ignore it since we will | |
1318 | * have already processed the original allocating I/O. | |
1319 | */ | |
1320 | if (flags & ZIO_FLAG_IO_ALLOCATING && | |
1321 | (vd != vd->vdev_top || (flags & ZIO_FLAG_IO_RETRY))) { | |
cc99f275 DB |
1322 | ASSERT(pio->io_metaslab_class != NULL); |
1323 | ASSERT(pio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
1324 | ASSERT(type == ZIO_TYPE_WRITE); |
1325 | ASSERT(priority == ZIO_PRIORITY_ASYNC_WRITE); | |
1326 | ASSERT(!(flags & ZIO_FLAG_IO_REPAIR)); | |
1327 | ASSERT(!(pio->io_flags & ZIO_FLAG_IO_REWRITE) || | |
1328 | pio->io_child_type == ZIO_CHILD_GANG); | |
1329 | ||
1330 | flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
1331 | } | |
1332 | ||
1333 | ||
2aa34383 | 1334 | zio = zio_create(pio, pio->io_spa, pio->io_txg, bp, data, size, size, |
428870ff BB |
1335 | done, private, type, priority, flags, vd, offset, &pio->io_bookmark, |
1336 | ZIO_STAGE_VDEV_IO_START >> 1, pipeline); | |
3dfb57a3 | 1337 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); |
34dc7c2f | 1338 | |
e8b96c60 MA |
1339 | zio->io_physdone = pio->io_physdone; |
1340 | if (vd->vdev_ops->vdev_op_leaf && zio->io_logical != NULL) | |
1341 | zio->io_logical->io_phys_children++; | |
1342 | ||
b128c09f | 1343 | return (zio); |
34dc7c2f BB |
1344 | } |
1345 | ||
b128c09f | 1346 | zio_t * |
a6255b7f | 1347 | zio_vdev_delegated_io(vdev_t *vd, uint64_t offset, abd_t *data, uint64_t size, |
9e052db4 | 1348 | zio_type_t type, zio_priority_t priority, enum zio_flag flags, |
e9aa730c | 1349 | zio_done_func_t *done, void *private) |
34dc7c2f | 1350 | { |
b128c09f | 1351 | zio_t *zio; |
34dc7c2f | 1352 | |
b128c09f | 1353 | ASSERT(vd->vdev_ops->vdev_op_leaf); |
34dc7c2f | 1354 | |
b128c09f | 1355 | zio = zio_create(NULL, vd->vdev_spa, 0, NULL, |
2aa34383 | 1356 | data, size, size, done, private, type, priority, |
e8b96c60 | 1357 | flags | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_RETRY | ZIO_FLAG_DELEGATED, |
b128c09f | 1358 | vd, offset, NULL, |
428870ff | 1359 | ZIO_STAGE_VDEV_IO_START >> 1, ZIO_VDEV_CHILD_PIPELINE); |
34dc7c2f | 1360 | |
b128c09f | 1361 | return (zio); |
34dc7c2f BB |
1362 | } |
1363 | ||
1364 | void | |
b128c09f | 1365 | zio_flush(zio_t *zio, vdev_t *vd) |
34dc7c2f | 1366 | { |
b128c09f | 1367 | zio_nowait(zio_ioctl(zio, zio->io_spa, vd, DKIOCFLUSHWRITECACHE, |
e8b96c60 | 1368 | NULL, NULL, |
b128c09f | 1369 | ZIO_FLAG_CANFAIL | ZIO_FLAG_DONT_PROPAGATE | ZIO_FLAG_DONT_RETRY)); |
34dc7c2f BB |
1370 | } |
1371 | ||
428870ff BB |
1372 | void |
1373 | zio_shrink(zio_t *zio, uint64_t size) | |
1374 | { | |
1ce23dca PS |
1375 | ASSERT3P(zio->io_executor, ==, NULL); |
1376 | ASSERT3U(zio->io_orig_size, ==, zio->io_size); | |
1377 | ASSERT3U(size, <=, zio->io_size); | |
428870ff BB |
1378 | |
1379 | /* | |
1380 | * We don't shrink for raidz because of problems with the | |
1381 | * reconstruction when reading back less than the block size. | |
1382 | * Note, BP_IS_RAIDZ() assumes no compression. | |
1383 | */ | |
1384 | ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF); | |
2aa34383 DK |
1385 | if (!BP_IS_RAIDZ(zio->io_bp)) { |
1386 | /* we are not doing a raw write */ | |
1387 | ASSERT3U(zio->io_size, ==, zio->io_lsize); | |
1388 | zio->io_orig_size = zio->io_size = zio->io_lsize = size; | |
1389 | } | |
428870ff BB |
1390 | } |
1391 | ||
34dc7c2f BB |
1392 | /* |
1393 | * ========================================================================== | |
b128c09f | 1394 | * Prepare to read and write logical blocks |
34dc7c2f BB |
1395 | * ========================================================================== |
1396 | */ | |
b128c09f | 1397 | |
62840030 | 1398 | static zio_t * |
b128c09f | 1399 | zio_read_bp_init(zio_t *zio) |
34dc7c2f | 1400 | { |
b128c09f | 1401 | blkptr_t *bp = zio->io_bp; |
b5256303 TC |
1402 | uint64_t psize = |
1403 | BP_IS_EMBEDDED(bp) ? BPE_GET_PSIZE(bp) : BP_GET_PSIZE(bp); | |
34dc7c2f | 1404 | |
a1d477c2 MA |
1405 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1406 | ||
fb5f0bc8 | 1407 | if (BP_GET_COMPRESS(bp) != ZIO_COMPRESS_OFF && |
9babb374 | 1408 | zio->io_child_type == ZIO_CHILD_LOGICAL && |
b5256303 | 1409 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { |
a6255b7f DQ |
1410 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), |
1411 | psize, psize, zio_decompress); | |
34dc7c2f | 1412 | } |
34dc7c2f | 1413 | |
b5256303 TC |
1414 | if (((BP_IS_PROTECTED(bp) && !(zio->io_flags & ZIO_FLAG_RAW_ENCRYPT)) || |
1415 | BP_HAS_INDIRECT_MAC_CKSUM(bp)) && | |
1416 | zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1417 | zio_push_transform(zio, abd_alloc_sametype(zio->io_abd, psize), | |
1418 | psize, psize, zio_decrypt); | |
1419 | } | |
1420 | ||
9b67f605 | 1421 | if (BP_IS_EMBEDDED(bp) && BPE_GET_ETYPE(bp) == BP_EMBEDDED_TYPE_DATA) { |
a6255b7f DQ |
1422 | int psize = BPE_GET_PSIZE(bp); |
1423 | void *data = abd_borrow_buf(zio->io_abd, psize); | |
1424 | ||
9b67f605 | 1425 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
a6255b7f DQ |
1426 | decode_embedded_bp_compressed(bp, data); |
1427 | abd_return_buf_copy(zio->io_abd, data, psize); | |
9b67f605 MA |
1428 | } else { |
1429 | ASSERT(!BP_IS_EMBEDDED(bp)); | |
a1d477c2 | 1430 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
9b67f605 MA |
1431 | } |
1432 | ||
9ae529ec | 1433 | if (!DMU_OT_IS_METADATA(BP_GET_TYPE(bp)) && BP_GET_LEVEL(bp) == 0) |
b128c09f BB |
1434 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; |
1435 | ||
428870ff BB |
1436 | if (BP_GET_TYPE(bp) == DMU_OT_DDT_ZAP) |
1437 | zio->io_flags |= ZIO_FLAG_DONT_CACHE; | |
1438 | ||
1439 | if (BP_GET_DEDUP(bp) && zio->io_child_type == ZIO_CHILD_LOGICAL) | |
1440 | zio->io_pipeline = ZIO_DDT_READ_PIPELINE; | |
1441 | ||
62840030 | 1442 | return (zio); |
34dc7c2f BB |
1443 | } |
1444 | ||
62840030 | 1445 | static zio_t * |
b128c09f | 1446 | zio_write_bp_init(zio_t *zio) |
34dc7c2f | 1447 | { |
b128c09f | 1448 | if (!IO_IS_ALLOCATING(zio)) |
62840030 | 1449 | return (zio); |
34dc7c2f | 1450 | |
428870ff BB |
1451 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1452 | ||
1453 | if (zio->io_bp_override) { | |
3dfb57a3 DB |
1454 | blkptr_t *bp = zio->io_bp; |
1455 | zio_prop_t *zp = &zio->io_prop; | |
1456 | ||
428870ff BB |
1457 | ASSERT(bp->blk_birth != zio->io_txg); |
1458 | ASSERT(BP_GET_DEDUP(zio->io_bp_override) == 0); | |
1459 | ||
1460 | *bp = *zio->io_bp_override; | |
1461 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1462 | ||
9b67f605 | 1463 | if (BP_IS_EMBEDDED(bp)) |
62840030 | 1464 | return (zio); |
9b67f605 | 1465 | |
03c6040b GW |
1466 | /* |
1467 | * If we've been overridden and nopwrite is set then | |
1468 | * set the flag accordingly to indicate that a nopwrite | |
1469 | * has already occurred. | |
1470 | */ | |
1471 | if (!BP_IS_HOLE(bp) && zp->zp_nopwrite) { | |
1472 | ASSERT(!zp->zp_dedup); | |
3dfb57a3 | 1473 | ASSERT3U(BP_GET_CHECKSUM(bp), ==, zp->zp_checksum); |
03c6040b | 1474 | zio->io_flags |= ZIO_FLAG_NOPWRITE; |
62840030 | 1475 | return (zio); |
03c6040b GW |
1476 | } |
1477 | ||
1478 | ASSERT(!zp->zp_nopwrite); | |
1479 | ||
428870ff | 1480 | if (BP_IS_HOLE(bp) || !zp->zp_dedup) |
62840030 | 1481 | return (zio); |
428870ff | 1482 | |
3c67d83a TH |
1483 | ASSERT((zio_checksum_table[zp->zp_checksum].ci_flags & |
1484 | ZCHECKSUM_FLAG_DEDUP) || zp->zp_dedup_verify); | |
428870ff | 1485 | |
b5256303 TC |
1486 | if (BP_GET_CHECKSUM(bp) == zp->zp_checksum && |
1487 | !zp->zp_encrypt) { | |
428870ff BB |
1488 | BP_SET_DEDUP(bp, 1); |
1489 | zio->io_pipeline |= ZIO_STAGE_DDT_WRITE; | |
62840030 | 1490 | return (zio); |
428870ff | 1491 | } |
3dfb57a3 DB |
1492 | |
1493 | /* | |
1494 | * We were unable to handle this as an override bp, treat | |
1495 | * it as a regular write I/O. | |
1496 | */ | |
5511754b | 1497 | zio->io_bp_override = NULL; |
3dfb57a3 DB |
1498 | *bp = zio->io_bp_orig; |
1499 | zio->io_pipeline = zio->io_orig_pipeline; | |
1500 | } | |
1501 | ||
62840030 | 1502 | return (zio); |
3dfb57a3 DB |
1503 | } |
1504 | ||
62840030 | 1505 | static zio_t * |
3dfb57a3 DB |
1506 | zio_write_compress(zio_t *zio) |
1507 | { | |
1508 | spa_t *spa = zio->io_spa; | |
1509 | zio_prop_t *zp = &zio->io_prop; | |
1510 | enum zio_compress compress = zp->zp_compress; | |
1511 | blkptr_t *bp = zio->io_bp; | |
1512 | uint64_t lsize = zio->io_lsize; | |
1513 | uint64_t psize = zio->io_size; | |
1514 | int pass = 1; | |
1515 | ||
3dfb57a3 DB |
1516 | /* |
1517 | * If our children haven't all reached the ready stage, | |
1518 | * wait for them and then repeat this pipeline stage. | |
1519 | */ | |
ddc751d5 GW |
1520 | if (zio_wait_for_children(zio, ZIO_CHILD_LOGICAL_BIT | |
1521 | ZIO_CHILD_GANG_BIT, ZIO_WAIT_READY)) { | |
62840030 | 1522 | return (NULL); |
ddc751d5 | 1523 | } |
3dfb57a3 DB |
1524 | |
1525 | if (!IO_IS_ALLOCATING(zio)) | |
62840030 | 1526 | return (zio); |
3dfb57a3 DB |
1527 | |
1528 | if (zio->io_children_ready != NULL) { | |
1529 | /* | |
1530 | * Now that all our children are ready, run the callback | |
1531 | * associated with this zio in case it wants to modify the | |
1532 | * data to be written. | |
1533 | */ | |
1534 | ASSERT3U(zp->zp_level, >, 0); | |
1535 | zio->io_children_ready(zio); | |
428870ff | 1536 | } |
34dc7c2f | 1537 | |
3dfb57a3 DB |
1538 | ASSERT(zio->io_child_type != ZIO_CHILD_DDT); |
1539 | ASSERT(zio->io_bp_override == NULL); | |
1540 | ||
b0bc7a84 | 1541 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg) { |
b128c09f BB |
1542 | /* |
1543 | * We're rewriting an existing block, which means we're | |
1544 | * working on behalf of spa_sync(). For spa_sync() to | |
1545 | * converge, it must eventually be the case that we don't | |
1546 | * have to allocate new blocks. But compression changes | |
1547 | * the blocksize, which forces a reallocate, and makes | |
1548 | * convergence take longer. Therefore, after the first | |
1549 | * few passes, stop compressing to ensure convergence. | |
1550 | */ | |
428870ff BB |
1551 | pass = spa_sync_pass(spa); |
1552 | ||
1553 | ASSERT(zio->io_txg == spa_syncing_txg(spa)); | |
1554 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1555 | ASSERT(!BP_GET_DEDUP(bp)); | |
34dc7c2f | 1556 | |
55d85d5a | 1557 | if (pass >= zfs_sync_pass_dont_compress) |
b128c09f | 1558 | compress = ZIO_COMPRESS_OFF; |
34dc7c2f | 1559 | |
b128c09f | 1560 | /* Make sure someone doesn't change their mind on overwrites */ |
9b67f605 | 1561 | ASSERT(BP_IS_EMBEDDED(bp) || MIN(zp->zp_copies + BP_IS_GANG(bp), |
428870ff | 1562 | spa_max_replication(spa)) == BP_GET_NDVAS(bp)); |
b128c09f | 1563 | } |
34dc7c2f | 1564 | |
2aa34383 | 1565 | /* If it's a compressed write that is not raw, compress the buffer. */ |
b5256303 TC |
1566 | if (compress != ZIO_COMPRESS_OFF && |
1567 | !(zio->io_flags & ZIO_FLAG_RAW_COMPRESS)) { | |
428870ff | 1568 | void *cbuf = zio_buf_alloc(lsize); |
a6255b7f | 1569 | psize = zio_compress_data(compress, zio->io_abd, cbuf, lsize); |
428870ff | 1570 | if (psize == 0 || psize == lsize) { |
b128c09f | 1571 | compress = ZIO_COMPRESS_OFF; |
428870ff | 1572 | zio_buf_free(cbuf, lsize); |
b5256303 TC |
1573 | } else if (!zp->zp_dedup && !zp->zp_encrypt && |
1574 | psize <= BPE_PAYLOAD_SIZE && | |
9b67f605 MA |
1575 | zp->zp_level == 0 && !DMU_OT_HAS_FILL(zp->zp_type) && |
1576 | spa_feature_is_enabled(spa, SPA_FEATURE_EMBEDDED_DATA)) { | |
1577 | encode_embedded_bp_compressed(bp, | |
1578 | cbuf, compress, lsize, psize); | |
1579 | BPE_SET_ETYPE(bp, BP_EMBEDDED_TYPE_DATA); | |
1580 | BP_SET_TYPE(bp, zio->io_prop.zp_type); | |
1581 | BP_SET_LEVEL(bp, zio->io_prop.zp_level); | |
1582 | zio_buf_free(cbuf, lsize); | |
1583 | bp->blk_birth = zio->io_txg; | |
1584 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
1585 | ASSERT(spa_feature_is_active(spa, | |
1586 | SPA_FEATURE_EMBEDDED_DATA)); | |
62840030 | 1587 | return (zio); |
428870ff | 1588 | } else { |
9b67f605 | 1589 | /* |
c3520e7f MA |
1590 | * Round up compressed size up to the ashift |
1591 | * of the smallest-ashift device, and zero the tail. | |
1592 | * This ensures that the compressed size of the BP | |
1593 | * (and thus compressratio property) are correct, | |
1594 | * in that we charge for the padding used to fill out | |
1595 | * the last sector. | |
9b67f605 | 1596 | */ |
c3520e7f | 1597 | ASSERT3U(spa->spa_min_ashift, >=, SPA_MINBLOCKSHIFT); |
1c27024e | 1598 | size_t rounded = (size_t)P2ROUNDUP(psize, |
c3520e7f MA |
1599 | 1ULL << spa->spa_min_ashift); |
1600 | if (rounded >= lsize) { | |
9b67f605 MA |
1601 | compress = ZIO_COMPRESS_OFF; |
1602 | zio_buf_free(cbuf, lsize); | |
c3520e7f | 1603 | psize = lsize; |
9b67f605 | 1604 | } else { |
a6255b7f DQ |
1605 | abd_t *cdata = abd_get_from_buf(cbuf, lsize); |
1606 | abd_take_ownership_of_buf(cdata, B_TRUE); | |
1607 | abd_zero_off(cdata, psize, rounded - psize); | |
c3520e7f | 1608 | psize = rounded; |
a6255b7f | 1609 | zio_push_transform(zio, cdata, |
9b67f605 MA |
1610 | psize, lsize, NULL); |
1611 | } | |
b128c09f | 1612 | } |
3dfb57a3 DB |
1613 | |
1614 | /* | |
1615 | * We were unable to handle this as an override bp, treat | |
1616 | * it as a regular write I/O. | |
1617 | */ | |
1618 | zio->io_bp_override = NULL; | |
1619 | *bp = zio->io_bp_orig; | |
1620 | zio->io_pipeline = zio->io_orig_pipeline; | |
1621 | ||
b1d21733 TC |
1622 | } else if ((zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) != 0 && |
1623 | zp->zp_type == DMU_OT_DNODE) { | |
1624 | /* | |
1625 | * The DMU actually relies on the zio layer's compression | |
1626 | * to free metadnode blocks that have had all contained | |
1627 | * dnodes freed. As a result, even when doing a raw | |
1628 | * receive, we must check whether the block can be compressed | |
1629 | * to a hole. | |
1630 | */ | |
1631 | psize = zio_compress_data(ZIO_COMPRESS_EMPTY, | |
1632 | zio->io_abd, NULL, lsize); | |
1633 | if (psize == 0) | |
1634 | compress = ZIO_COMPRESS_OFF; | |
2aa34383 DK |
1635 | } else { |
1636 | ASSERT3U(psize, !=, 0); | |
b128c09f | 1637 | } |
34dc7c2f | 1638 | |
b128c09f BB |
1639 | /* |
1640 | * The final pass of spa_sync() must be all rewrites, but the first | |
1641 | * few passes offer a trade-off: allocating blocks defers convergence, | |
1642 | * but newly allocated blocks are sequential, so they can be written | |
1643 | * to disk faster. Therefore, we allow the first few passes of | |
1644 | * spa_sync() to allocate new blocks, but force rewrites after that. | |
1645 | * There should only be a handful of blocks after pass 1 in any case. | |
1646 | */ | |
b0bc7a84 MG |
1647 | if (!BP_IS_HOLE(bp) && bp->blk_birth == zio->io_txg && |
1648 | BP_GET_PSIZE(bp) == psize && | |
55d85d5a | 1649 | pass >= zfs_sync_pass_rewrite) { |
cc99f275 | 1650 | VERIFY3U(psize, !=, 0); |
1c27024e | 1651 | enum zio_stage gang_stages = zio->io_pipeline & ZIO_GANG_STAGES; |
cc99f275 | 1652 | |
b128c09f BB |
1653 | zio->io_pipeline = ZIO_REWRITE_PIPELINE | gang_stages; |
1654 | zio->io_flags |= ZIO_FLAG_IO_REWRITE; | |
1655 | } else { | |
1656 | BP_ZERO(bp); | |
1657 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
1658 | } | |
34dc7c2f | 1659 | |
428870ff | 1660 | if (psize == 0) { |
b0bc7a84 MG |
1661 | if (zio->io_bp_orig.blk_birth != 0 && |
1662 | spa_feature_is_active(spa, SPA_FEATURE_HOLE_BIRTH)) { | |
1663 | BP_SET_LSIZE(bp, lsize); | |
1664 | BP_SET_TYPE(bp, zp->zp_type); | |
1665 | BP_SET_LEVEL(bp, zp->zp_level); | |
1666 | BP_SET_BIRTH(bp, zio->io_txg, 0); | |
1667 | } | |
b128c09f BB |
1668 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
1669 | } else { | |
1670 | ASSERT(zp->zp_checksum != ZIO_CHECKSUM_GANG_HEADER); | |
1671 | BP_SET_LSIZE(bp, lsize); | |
b0bc7a84 MG |
1672 | BP_SET_TYPE(bp, zp->zp_type); |
1673 | BP_SET_LEVEL(bp, zp->zp_level); | |
428870ff | 1674 | BP_SET_PSIZE(bp, psize); |
b128c09f BB |
1675 | BP_SET_COMPRESS(bp, compress); |
1676 | BP_SET_CHECKSUM(bp, zp->zp_checksum); | |
428870ff | 1677 | BP_SET_DEDUP(bp, zp->zp_dedup); |
b128c09f | 1678 | BP_SET_BYTEORDER(bp, ZFS_HOST_BYTEORDER); |
428870ff BB |
1679 | if (zp->zp_dedup) { |
1680 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1681 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
b5256303 TC |
1682 | ASSERT(!zp->zp_encrypt || |
1683 | DMU_OT_IS_ENCRYPTED(zp->zp_type)); | |
428870ff BB |
1684 | zio->io_pipeline = ZIO_DDT_WRITE_PIPELINE; |
1685 | } | |
03c6040b GW |
1686 | if (zp->zp_nopwrite) { |
1687 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
1688 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
1689 | zio->io_pipeline |= ZIO_STAGE_NOP_WRITE; | |
1690 | } | |
428870ff | 1691 | } |
62840030 | 1692 | return (zio); |
428870ff BB |
1693 | } |
1694 | ||
62840030 | 1695 | static zio_t * |
428870ff BB |
1696 | zio_free_bp_init(zio_t *zio) |
1697 | { | |
1698 | blkptr_t *bp = zio->io_bp; | |
1699 | ||
1700 | if (zio->io_child_type == ZIO_CHILD_LOGICAL) { | |
1701 | if (BP_GET_DEDUP(bp)) | |
1702 | zio->io_pipeline = ZIO_DDT_FREE_PIPELINE; | |
b128c09f | 1703 | } |
34dc7c2f | 1704 | |
a1d477c2 MA |
1705 | ASSERT3P(zio->io_bp, ==, &zio->io_bp_copy); |
1706 | ||
62840030 | 1707 | return (zio); |
34dc7c2f BB |
1708 | } |
1709 | ||
b128c09f BB |
1710 | /* |
1711 | * ========================================================================== | |
1712 | * Execute the I/O pipeline | |
1713 | * ========================================================================== | |
1714 | */ | |
1715 | ||
1716 | static void | |
7ef5e54e | 1717 | zio_taskq_dispatch(zio_t *zio, zio_taskq_type_t q, boolean_t cutinline) |
34dc7c2f | 1718 | { |
428870ff | 1719 | spa_t *spa = zio->io_spa; |
b128c09f | 1720 | zio_type_t t = zio->io_type; |
a38718a6 | 1721 | int flags = (cutinline ? TQ_FRONT : 0); |
34dc7c2f BB |
1722 | |
1723 | /* | |
9babb374 BB |
1724 | * If we're a config writer or a probe, the normal issue and |
1725 | * interrupt threads may all be blocked waiting for the config lock. | |
1726 | * In this case, select the otherwise-unused taskq for ZIO_TYPE_NULL. | |
34dc7c2f | 1727 | */ |
9babb374 | 1728 | if (zio->io_flags & (ZIO_FLAG_CONFIG_WRITER | ZIO_FLAG_PROBE)) |
b128c09f | 1729 | t = ZIO_TYPE_NULL; |
34dc7c2f BB |
1730 | |
1731 | /* | |
b128c09f | 1732 | * A similar issue exists for the L2ARC write thread until L2ARC 2.0. |
34dc7c2f | 1733 | */ |
b128c09f BB |
1734 | if (t == ZIO_TYPE_WRITE && zio->io_vd && zio->io_vd->vdev_aux) |
1735 | t = ZIO_TYPE_NULL; | |
34dc7c2f | 1736 | |
428870ff | 1737 | /* |
7ef5e54e AL |
1738 | * If this is a high priority I/O, then use the high priority taskq if |
1739 | * available. | |
428870ff | 1740 | */ |
18b14b17 GW |
1741 | if ((zio->io_priority == ZIO_PRIORITY_NOW || |
1742 | zio->io_priority == ZIO_PRIORITY_SYNC_WRITE) && | |
7ef5e54e | 1743 | spa->spa_zio_taskq[t][q + 1].stqs_count != 0) |
428870ff BB |
1744 | q++; |
1745 | ||
1746 | ASSERT3U(q, <, ZIO_TASKQ_TYPES); | |
5cc556b4 | 1747 | |
a38718a6 GA |
1748 | /* |
1749 | * NB: We are assuming that the zio can only be dispatched | |
1750 | * to a single taskq at a time. It would be a grievous error | |
1751 | * to dispatch the zio to another taskq at the same time. | |
1752 | */ | |
1753 | ASSERT(taskq_empty_ent(&zio->io_tqent)); | |
7ef5e54e AL |
1754 | spa_taskq_dispatch_ent(spa, t, q, (task_func_t *)zio_execute, zio, |
1755 | flags, &zio->io_tqent); | |
b128c09f | 1756 | } |
34dc7c2f | 1757 | |
b128c09f | 1758 | static boolean_t |
7ef5e54e | 1759 | zio_taskq_member(zio_t *zio, zio_taskq_type_t q) |
b128c09f BB |
1760 | { |
1761 | kthread_t *executor = zio->io_executor; | |
1762 | spa_t *spa = zio->io_spa; | |
34dc7c2f | 1763 | |
1c27024e | 1764 | for (zio_type_t t = 0; t < ZIO_TYPES; t++) { |
7ef5e54e AL |
1765 | spa_taskqs_t *tqs = &spa->spa_zio_taskq[t][q]; |
1766 | uint_t i; | |
1767 | for (i = 0; i < tqs->stqs_count; i++) { | |
1768 | if (taskq_member(tqs->stqs_taskq[i], executor)) | |
1769 | return (B_TRUE); | |
1770 | } | |
1771 | } | |
34dc7c2f | 1772 | |
b128c09f BB |
1773 | return (B_FALSE); |
1774 | } | |
34dc7c2f | 1775 | |
62840030 | 1776 | static zio_t * |
b128c09f BB |
1777 | zio_issue_async(zio_t *zio) |
1778 | { | |
428870ff | 1779 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); |
b128c09f | 1780 | |
62840030 | 1781 | return (NULL); |
34dc7c2f BB |
1782 | } |
1783 | ||
b128c09f BB |
1784 | void |
1785 | zio_interrupt(zio_t *zio) | |
34dc7c2f | 1786 | { |
428870ff | 1787 | zio_taskq_dispatch(zio, ZIO_TASKQ_INTERRUPT, B_FALSE); |
b128c09f | 1788 | } |
34dc7c2f | 1789 | |
26ef0cc7 TH |
1790 | void |
1791 | zio_delay_interrupt(zio_t *zio) | |
1792 | { | |
1793 | /* | |
1794 | * The timeout_generic() function isn't defined in userspace, so | |
1795 | * rather than trying to implement the function, the zio delay | |
1796 | * functionality has been disabled for userspace builds. | |
1797 | */ | |
1798 | ||
1799 | #ifdef _KERNEL | |
1800 | /* | |
1801 | * If io_target_timestamp is zero, then no delay has been registered | |
1802 | * for this IO, thus jump to the end of this function and "skip" the | |
1803 | * delay; issuing it directly to the zio layer. | |
1804 | */ | |
1805 | if (zio->io_target_timestamp != 0) { | |
1806 | hrtime_t now = gethrtime(); | |
1807 | ||
1808 | if (now >= zio->io_target_timestamp) { | |
1809 | /* | |
1810 | * This IO has already taken longer than the target | |
1811 | * delay to complete, so we don't want to delay it | |
1812 | * any longer; we "miss" the delay and issue it | |
1813 | * directly to the zio layer. This is likely due to | |
1814 | * the target latency being set to a value less than | |
1815 | * the underlying hardware can satisfy (e.g. delay | |
1816 | * set to 1ms, but the disks take 10ms to complete an | |
1817 | * IO request). | |
1818 | */ | |
1819 | ||
1820 | DTRACE_PROBE2(zio__delay__miss, zio_t *, zio, | |
1821 | hrtime_t, now); | |
1822 | ||
1823 | zio_interrupt(zio); | |
1824 | } else { | |
1825 | taskqid_t tid; | |
1826 | hrtime_t diff = zio->io_target_timestamp - now; | |
1827 | clock_t expire_at_tick = ddi_get_lbolt() + | |
1828 | NSEC_TO_TICK(diff); | |
1829 | ||
1830 | DTRACE_PROBE3(zio__delay__hit, zio_t *, zio, | |
1831 | hrtime_t, now, hrtime_t, diff); | |
1832 | ||
1833 | if (NSEC_TO_TICK(diff) == 0) { | |
1834 | /* Our delay is less than a jiffy - just spin */ | |
1835 | zfs_sleep_until(zio->io_target_timestamp); | |
1836 | } else { | |
1837 | /* | |
1838 | * Use taskq_dispatch_delay() in the place of | |
1839 | * OpenZFS's timeout_generic(). | |
1840 | */ | |
1841 | tid = taskq_dispatch_delay(system_taskq, | |
02730c33 | 1842 | (task_func_t *)zio_interrupt, |
26ef0cc7 | 1843 | zio, TQ_NOSLEEP, expire_at_tick); |
48d3eb40 | 1844 | if (tid == TASKQID_INVALID) { |
26ef0cc7 TH |
1845 | /* |
1846 | * Couldn't allocate a task. Just | |
1847 | * finish the zio without a delay. | |
1848 | */ | |
1849 | zio_interrupt(zio); | |
1850 | } | |
1851 | } | |
1852 | } | |
1853 | return; | |
1854 | } | |
1855 | #endif | |
1856 | DTRACE_PROBE1(zio__delay__skip, zio_t *, zio); | |
1857 | zio_interrupt(zio); | |
1858 | } | |
1859 | ||
8fb1ede1 BB |
1860 | static void |
1861 | zio_deadman_impl(zio_t *pio) | |
1862 | { | |
1863 | zio_t *cio, *cio_next; | |
1864 | zio_link_t *zl = NULL; | |
1865 | vdev_t *vd = pio->io_vd; | |
1866 | ||
1867 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { | |
1868 | vdev_queue_t *vq = &vd->vdev_queue; | |
1869 | zbookmark_phys_t *zb = &pio->io_bookmark; | |
1870 | uint64_t delta = gethrtime() - pio->io_timestamp; | |
1871 | uint64_t failmode = spa_get_deadman_failmode(pio->io_spa); | |
1872 | ||
1873 | zfs_dbgmsg("slow zio: zio=%p timestamp=%llu " | |
1874 | "delta=%llu queued=%llu io=%llu " | |
1875 | "path=%s last=%llu " | |
1876 | "type=%d priority=%d flags=0x%x " | |
1877 | "stage=0x%x pipeline=0x%x pipeline-trace=0x%x " | |
1878 | "objset=%llu object=%llu level=%llu blkid=%llu " | |
1879 | "offset=%llu size=%llu error=%d", | |
1880 | pio, pio->io_timestamp, | |
1881 | delta, pio->io_delta, pio->io_delay, | |
1882 | vd->vdev_path, vq->vq_io_complete_ts, | |
1883 | pio->io_type, pio->io_priority, pio->io_flags, | |
1884 | pio->io_state, pio->io_pipeline, pio->io_pipeline_trace, | |
1885 | zb->zb_objset, zb->zb_object, zb->zb_level, zb->zb_blkid, | |
1886 | pio->io_offset, pio->io_size, pio->io_error); | |
1887 | zfs_ereport_post(FM_EREPORT_ZFS_DEADMAN, | |
1888 | pio->io_spa, vd, zb, pio, 0, 0); | |
1889 | ||
1890 | if (failmode == ZIO_FAILURE_MODE_CONTINUE && | |
1891 | taskq_empty_ent(&pio->io_tqent)) { | |
1892 | zio_interrupt(pio); | |
1893 | } | |
1894 | } | |
1895 | ||
1896 | mutex_enter(&pio->io_lock); | |
1897 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
1898 | cio_next = zio_walk_children(pio, &zl); | |
1899 | zio_deadman_impl(cio); | |
1900 | } | |
1901 | mutex_exit(&pio->io_lock); | |
1902 | } | |
1903 | ||
1904 | /* | |
1905 | * Log the critical information describing this zio and all of its children | |
1906 | * using the zfs_dbgmsg() interface then post deadman event for the ZED. | |
1907 | */ | |
1908 | void | |
1909 | zio_deadman(zio_t *pio, char *tag) | |
1910 | { | |
1911 | spa_t *spa = pio->io_spa; | |
1912 | char *name = spa_name(spa); | |
1913 | ||
1914 | if (!zfs_deadman_enabled || spa_suspended(spa)) | |
1915 | return; | |
1916 | ||
1917 | zio_deadman_impl(pio); | |
1918 | ||
1919 | switch (spa_get_deadman_failmode(spa)) { | |
1920 | case ZIO_FAILURE_MODE_WAIT: | |
1921 | zfs_dbgmsg("%s waiting for hung I/O to pool '%s'", tag, name); | |
1922 | break; | |
1923 | ||
1924 | case ZIO_FAILURE_MODE_CONTINUE: | |
1925 | zfs_dbgmsg("%s restarting hung I/O for pool '%s'", tag, name); | |
1926 | break; | |
1927 | ||
1928 | case ZIO_FAILURE_MODE_PANIC: | |
1929 | fm_panic("%s determined I/O to pool '%s' is hung.", tag, name); | |
1930 | break; | |
1931 | } | |
1932 | } | |
1933 | ||
b128c09f BB |
1934 | /* |
1935 | * Execute the I/O pipeline until one of the following occurs: | |
1936 | * (1) the I/O completes; (2) the pipeline stalls waiting for | |
1937 | * dependent child I/Os; (3) the I/O issues, so we're waiting | |
1938 | * for an I/O completion interrupt; (4) the I/O is delegated by | |
1939 | * vdev-level caching or aggregation; (5) the I/O is deferred | |
1940 | * due to vdev-level queueing; (6) the I/O is handed off to | |
1941 | * another thread. In all cases, the pipeline stops whenever | |
8e07b99b | 1942 | * there's no CPU work; it never burns a thread in cv_wait_io(). |
b128c09f BB |
1943 | * |
1944 | * There's no locking on io_stage because there's no legitimate way | |
1945 | * for multiple threads to be attempting to process the same I/O. | |
1946 | */ | |
428870ff | 1947 | static zio_pipe_stage_t *zio_pipeline[]; |
34dc7c2f | 1948 | |
da6b4005 NB |
1949 | /* |
1950 | * zio_execute() is a wrapper around the static function | |
1951 | * __zio_execute() so that we can force __zio_execute() to be | |
1952 | * inlined. This reduces stack overhead which is important | |
1953 | * because __zio_execute() is called recursively in several zio | |
1954 | * code paths. zio_execute() itself cannot be inlined because | |
1955 | * it is externally visible. | |
1956 | */ | |
b128c09f BB |
1957 | void |
1958 | zio_execute(zio_t *zio) | |
da6b4005 | 1959 | { |
92119cc2 BB |
1960 | fstrans_cookie_t cookie; |
1961 | ||
1962 | cookie = spl_fstrans_mark(); | |
da6b4005 | 1963 | __zio_execute(zio); |
92119cc2 | 1964 | spl_fstrans_unmark(cookie); |
da6b4005 NB |
1965 | } |
1966 | ||
b58986ee BB |
1967 | /* |
1968 | * Used to determine if in the current context the stack is sized large | |
1969 | * enough to allow zio_execute() to be called recursively. A minimum | |
1970 | * stack size of 16K is required to avoid needing to re-dispatch the zio. | |
1971 | */ | |
1972 | boolean_t | |
1973 | zio_execute_stack_check(zio_t *zio) | |
1974 | { | |
1975 | #if !defined(HAVE_LARGE_STACKS) | |
1976 | dsl_pool_t *dp = spa_get_dsl(zio->io_spa); | |
1977 | ||
1978 | /* Executing in txg_sync_thread() context. */ | |
1979 | if (dp && curthread == dp->dp_tx.tx_sync_thread) | |
1980 | return (B_TRUE); | |
1981 | ||
1982 | /* Pool initialization outside of zio_taskq context. */ | |
1983 | if (dp && spa_is_initializing(dp->dp_spa) && | |
1984 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE) && | |
1985 | !zio_taskq_member(zio, ZIO_TASKQ_ISSUE_HIGH)) | |
1986 | return (B_TRUE); | |
1987 | #endif /* HAVE_LARGE_STACKS */ | |
1988 | ||
1989 | return (B_FALSE); | |
1990 | } | |
1991 | ||
da6b4005 NB |
1992 | __attribute__((always_inline)) |
1993 | static inline void | |
1994 | __zio_execute(zio_t *zio) | |
b128c09f | 1995 | { |
3dfb57a3 DB |
1996 | ASSERT3U(zio->io_queued_timestamp, >, 0); |
1997 | ||
b128c09f | 1998 | while (zio->io_stage < ZIO_STAGE_DONE) { |
428870ff BB |
1999 | enum zio_stage pipeline = zio->io_pipeline; |
2000 | enum zio_stage stage = zio->io_stage; | |
62840030 MA |
2001 | |
2002 | zio->io_executor = curthread; | |
34dc7c2f | 2003 | |
b128c09f | 2004 | ASSERT(!MUTEX_HELD(&zio->io_lock)); |
428870ff BB |
2005 | ASSERT(ISP2(stage)); |
2006 | ASSERT(zio->io_stall == NULL); | |
34dc7c2f | 2007 | |
428870ff BB |
2008 | do { |
2009 | stage <<= 1; | |
2010 | } while ((stage & pipeline) == 0); | |
b128c09f BB |
2011 | |
2012 | ASSERT(stage <= ZIO_STAGE_DONE); | |
34dc7c2f BB |
2013 | |
2014 | /* | |
b128c09f BB |
2015 | * If we are in interrupt context and this pipeline stage |
2016 | * will grab a config lock that is held across I/O, | |
428870ff BB |
2017 | * or may wait for an I/O that needs an interrupt thread |
2018 | * to complete, issue async to avoid deadlock. | |
2019 | * | |
2020 | * For VDEV_IO_START, we cut in line so that the io will | |
2021 | * be sent to disk promptly. | |
34dc7c2f | 2022 | */ |
91579709 BB |
2023 | if ((stage & ZIO_BLOCKING_STAGES) && zio->io_vd == NULL && |
2024 | zio_taskq_member(zio, ZIO_TASKQ_INTERRUPT)) { | |
b58986ee BB |
2025 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? |
2026 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
91579709 BB |
2027 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
2028 | return; | |
2029 | } | |
2030 | ||
2031 | /* | |
b58986ee BB |
2032 | * If the current context doesn't have large enough stacks |
2033 | * the zio must be issued asynchronously to prevent overflow. | |
91579709 | 2034 | */ |
b58986ee BB |
2035 | if (zio_execute_stack_check(zio)) { |
2036 | boolean_t cut = (stage == ZIO_STAGE_VDEV_IO_START) ? | |
2037 | zio_requeue_io_start_cut_in_line : B_FALSE; | |
428870ff | 2038 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, cut); |
b128c09f | 2039 | return; |
34dc7c2f BB |
2040 | } |
2041 | ||
b128c09f | 2042 | zio->io_stage = stage; |
3dfb57a3 | 2043 | zio->io_pipeline_trace |= zio->io_stage; |
34dc7c2f | 2044 | |
62840030 MA |
2045 | /* |
2046 | * The zio pipeline stage returns the next zio to execute | |
2047 | * (typically the same as this one), or NULL if we should | |
2048 | * stop. | |
2049 | */ | |
2050 | zio = zio_pipeline[highbit64(stage) - 1](zio); | |
34dc7c2f | 2051 | |
62840030 MA |
2052 | if (zio == NULL) |
2053 | return; | |
b128c09f | 2054 | } |
34dc7c2f BB |
2055 | } |
2056 | ||
da6b4005 | 2057 | |
b128c09f BB |
2058 | /* |
2059 | * ========================================================================== | |
2060 | * Initiate I/O, either sync or async | |
2061 | * ========================================================================== | |
2062 | */ | |
2063 | int | |
2064 | zio_wait(zio_t *zio) | |
34dc7c2f | 2065 | { |
8fb1ede1 | 2066 | long timeout = MSEC_TO_TICK(zfs_deadman_ziotime_ms); |
b128c09f | 2067 | int error; |
34dc7c2f | 2068 | |
1ce23dca PS |
2069 | ASSERT3S(zio->io_stage, ==, ZIO_STAGE_OPEN); |
2070 | ASSERT3P(zio->io_executor, ==, NULL); | |
34dc7c2f | 2071 | |
b128c09f | 2072 | zio->io_waiter = curthread; |
3dfb57a3 DB |
2073 | ASSERT0(zio->io_queued_timestamp); |
2074 | zio->io_queued_timestamp = gethrtime(); | |
34dc7c2f | 2075 | |
da6b4005 | 2076 | __zio_execute(zio); |
34dc7c2f | 2077 | |
b128c09f | 2078 | mutex_enter(&zio->io_lock); |
8fb1ede1 BB |
2079 | while (zio->io_executor != NULL) { |
2080 | error = cv_timedwait_io(&zio->io_cv, &zio->io_lock, | |
2081 | ddi_get_lbolt() + timeout); | |
2082 | ||
2083 | if (zfs_deadman_enabled && error == -1 && | |
2084 | gethrtime() - zio->io_queued_timestamp > | |
2085 | spa_deadman_ziotime(zio->io_spa)) { | |
2086 | mutex_exit(&zio->io_lock); | |
2087 | timeout = MSEC_TO_TICK(zfs_deadman_checktime_ms); | |
2088 | zio_deadman(zio, FTAG); | |
2089 | mutex_enter(&zio->io_lock); | |
2090 | } | |
2091 | } | |
b128c09f | 2092 | mutex_exit(&zio->io_lock); |
34dc7c2f | 2093 | |
b128c09f BB |
2094 | error = zio->io_error; |
2095 | zio_destroy(zio); | |
34dc7c2f | 2096 | |
b128c09f BB |
2097 | return (error); |
2098 | } | |
34dc7c2f | 2099 | |
b128c09f BB |
2100 | void |
2101 | zio_nowait(zio_t *zio) | |
2102 | { | |
1ce23dca | 2103 | ASSERT3P(zio->io_executor, ==, NULL); |
34dc7c2f | 2104 | |
d164b209 BB |
2105 | if (zio->io_child_type == ZIO_CHILD_LOGICAL && |
2106 | zio_unique_parent(zio) == NULL) { | |
8878261f BB |
2107 | zio_t *pio; |
2108 | ||
34dc7c2f | 2109 | /* |
b128c09f | 2110 | * This is a logical async I/O with no parent to wait for it. |
9babb374 BB |
2111 | * We add it to the spa_async_root_zio "Godfather" I/O which |
2112 | * will ensure they complete prior to unloading the pool. | |
34dc7c2f | 2113 | */ |
b128c09f | 2114 | spa_t *spa = zio->io_spa; |
8878261f BB |
2115 | kpreempt_disable(); |
2116 | pio = spa->spa_async_zio_root[CPU_SEQID]; | |
2117 | kpreempt_enable(); | |
9babb374 | 2118 | |
8878261f | 2119 | zio_add_child(pio, zio); |
b128c09f | 2120 | } |
34dc7c2f | 2121 | |
3dfb57a3 DB |
2122 | ASSERT0(zio->io_queued_timestamp); |
2123 | zio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2124 | __zio_execute(zio); |
b128c09f | 2125 | } |
34dc7c2f | 2126 | |
b128c09f BB |
2127 | /* |
2128 | * ========================================================================== | |
1ce23dca | 2129 | * Reexecute, cancel, or suspend/resume failed I/O |
b128c09f BB |
2130 | * ========================================================================== |
2131 | */ | |
34dc7c2f | 2132 | |
b128c09f BB |
2133 | static void |
2134 | zio_reexecute(zio_t *pio) | |
2135 | { | |
d164b209 BB |
2136 | zio_t *cio, *cio_next; |
2137 | ||
2138 | ASSERT(pio->io_child_type == ZIO_CHILD_LOGICAL); | |
2139 | ASSERT(pio->io_orig_stage == ZIO_STAGE_OPEN); | |
9babb374 BB |
2140 | ASSERT(pio->io_gang_leader == NULL); |
2141 | ASSERT(pio->io_gang_tree == NULL); | |
34dc7c2f | 2142 | |
b128c09f BB |
2143 | pio->io_flags = pio->io_orig_flags; |
2144 | pio->io_stage = pio->io_orig_stage; | |
2145 | pio->io_pipeline = pio->io_orig_pipeline; | |
2146 | pio->io_reexecute = 0; | |
03c6040b | 2147 | pio->io_flags |= ZIO_FLAG_REEXECUTED; |
3dfb57a3 | 2148 | pio->io_pipeline_trace = 0; |
b128c09f | 2149 | pio->io_error = 0; |
1c27024e | 2150 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2151 | pio->io_state[w] = 0; |
1c27024e | 2152 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
b128c09f | 2153 | pio->io_child_error[c] = 0; |
34dc7c2f | 2154 | |
428870ff BB |
2155 | if (IO_IS_ALLOCATING(pio)) |
2156 | BP_ZERO(pio->io_bp); | |
34dc7c2f | 2157 | |
b128c09f BB |
2158 | /* |
2159 | * As we reexecute pio's children, new children could be created. | |
d164b209 | 2160 | * New children go to the head of pio's io_child_list, however, |
b128c09f | 2161 | * so we will (correctly) not reexecute them. The key is that |
d164b209 BB |
2162 | * the remainder of pio's io_child_list, from 'cio_next' onward, |
2163 | * cannot be affected by any side effects of reexecuting 'cio'. | |
b128c09f | 2164 | */ |
1c27024e | 2165 | zio_link_t *zl = NULL; |
a8b2e306 | 2166 | mutex_enter(&pio->io_lock); |
3dfb57a3 DB |
2167 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { |
2168 | cio_next = zio_walk_children(pio, &zl); | |
1c27024e | 2169 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) |
d164b209 | 2170 | pio->io_children[cio->io_child_type][w]++; |
b128c09f | 2171 | mutex_exit(&pio->io_lock); |
d164b209 | 2172 | zio_reexecute(cio); |
a8b2e306 | 2173 | mutex_enter(&pio->io_lock); |
34dc7c2f | 2174 | } |
a8b2e306 | 2175 | mutex_exit(&pio->io_lock); |
34dc7c2f | 2176 | |
b128c09f BB |
2177 | /* |
2178 | * Now that all children have been reexecuted, execute the parent. | |
9babb374 | 2179 | * We don't reexecute "The Godfather" I/O here as it's the |
9e2c3bb4 | 2180 | * responsibility of the caller to wait on it. |
b128c09f | 2181 | */ |
3dfb57a3 DB |
2182 | if (!(pio->io_flags & ZIO_FLAG_GODFATHER)) { |
2183 | pio->io_queued_timestamp = gethrtime(); | |
da6b4005 | 2184 | __zio_execute(pio); |
3dfb57a3 | 2185 | } |
34dc7c2f BB |
2186 | } |
2187 | ||
b128c09f | 2188 | void |
cec3a0a1 | 2189 | zio_suspend(spa_t *spa, zio_t *zio, zio_suspend_reason_t reason) |
34dc7c2f | 2190 | { |
b128c09f BB |
2191 | if (spa_get_failmode(spa) == ZIO_FAILURE_MODE_PANIC) |
2192 | fm_panic("Pool '%s' has encountered an uncorrectable I/O " | |
2193 | "failure and the failure mode property for this pool " | |
2194 | "is set to panic.", spa_name(spa)); | |
34dc7c2f | 2195 | |
bf89c199 BB |
2196 | cmn_err(CE_WARN, "Pool '%s' has encountered an uncorrectable I/O " |
2197 | "failure and has been suspended.\n", spa_name(spa)); | |
2198 | ||
b5256303 TC |
2199 | zfs_ereport_post(FM_EREPORT_ZFS_IO_FAILURE, spa, NULL, |
2200 | NULL, NULL, 0, 0); | |
34dc7c2f | 2201 | |
b128c09f | 2202 | mutex_enter(&spa->spa_suspend_lock); |
34dc7c2f | 2203 | |
b128c09f | 2204 | if (spa->spa_suspend_zio_root == NULL) |
9babb374 BB |
2205 | spa->spa_suspend_zio_root = zio_root(spa, NULL, NULL, |
2206 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2207 | ZIO_FLAG_GODFATHER); | |
34dc7c2f | 2208 | |
cec3a0a1 | 2209 | spa->spa_suspended = reason; |
34dc7c2f | 2210 | |
b128c09f | 2211 | if (zio != NULL) { |
9babb374 | 2212 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f BB |
2213 | ASSERT(zio != spa->spa_suspend_zio_root); |
2214 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
d164b209 | 2215 | ASSERT(zio_unique_parent(zio) == NULL); |
b128c09f BB |
2216 | ASSERT(zio->io_stage == ZIO_STAGE_DONE); |
2217 | zio_add_child(spa->spa_suspend_zio_root, zio); | |
2218 | } | |
34dc7c2f | 2219 | |
b128c09f BB |
2220 | mutex_exit(&spa->spa_suspend_lock); |
2221 | } | |
34dc7c2f | 2222 | |
9babb374 | 2223 | int |
b128c09f BB |
2224 | zio_resume(spa_t *spa) |
2225 | { | |
9babb374 | 2226 | zio_t *pio; |
34dc7c2f BB |
2227 | |
2228 | /* | |
b128c09f | 2229 | * Reexecute all previously suspended i/o. |
34dc7c2f | 2230 | */ |
b128c09f | 2231 | mutex_enter(&spa->spa_suspend_lock); |
cec3a0a1 | 2232 | spa->spa_suspended = ZIO_SUSPEND_NONE; |
b128c09f BB |
2233 | cv_broadcast(&spa->spa_suspend_cv); |
2234 | pio = spa->spa_suspend_zio_root; | |
2235 | spa->spa_suspend_zio_root = NULL; | |
2236 | mutex_exit(&spa->spa_suspend_lock); | |
2237 | ||
2238 | if (pio == NULL) | |
9babb374 | 2239 | return (0); |
34dc7c2f | 2240 | |
9babb374 BB |
2241 | zio_reexecute(pio); |
2242 | return (zio_wait(pio)); | |
b128c09f BB |
2243 | } |
2244 | ||
2245 | void | |
2246 | zio_resume_wait(spa_t *spa) | |
2247 | { | |
2248 | mutex_enter(&spa->spa_suspend_lock); | |
2249 | while (spa_suspended(spa)) | |
2250 | cv_wait(&spa->spa_suspend_cv, &spa->spa_suspend_lock); | |
2251 | mutex_exit(&spa->spa_suspend_lock); | |
34dc7c2f BB |
2252 | } |
2253 | ||
2254 | /* | |
2255 | * ========================================================================== | |
b128c09f BB |
2256 | * Gang blocks. |
2257 | * | |
2258 | * A gang block is a collection of small blocks that looks to the DMU | |
2259 | * like one large block. When zio_dva_allocate() cannot find a block | |
2260 | * of the requested size, due to either severe fragmentation or the pool | |
2261 | * being nearly full, it calls zio_write_gang_block() to construct the | |
2262 | * block from smaller fragments. | |
2263 | * | |
2264 | * A gang block consists of a gang header (zio_gbh_phys_t) and up to | |
2265 | * three (SPA_GBH_NBLKPTRS) gang members. The gang header is just like | |
2266 | * an indirect block: it's an array of block pointers. It consumes | |
2267 | * only one sector and hence is allocatable regardless of fragmentation. | |
2268 | * The gang header's bps point to its gang members, which hold the data. | |
2269 | * | |
2270 | * Gang blocks are self-checksumming, using the bp's <vdev, offset, txg> | |
2271 | * as the verifier to ensure uniqueness of the SHA256 checksum. | |
2272 | * Critically, the gang block bp's blk_cksum is the checksum of the data, | |
2273 | * not the gang header. This ensures that data block signatures (needed for | |
2274 | * deduplication) are independent of how the block is physically stored. | |
2275 | * | |
2276 | * Gang blocks can be nested: a gang member may itself be a gang block. | |
2277 | * Thus every gang block is a tree in which root and all interior nodes are | |
2278 | * gang headers, and the leaves are normal blocks that contain user data. | |
2279 | * The root of the gang tree is called the gang leader. | |
2280 | * | |
2281 | * To perform any operation (read, rewrite, free, claim) on a gang block, | |
2282 | * zio_gang_assemble() first assembles the gang tree (minus data leaves) | |
2283 | * in the io_gang_tree field of the original logical i/o by recursively | |
2284 | * reading the gang leader and all gang headers below it. This yields | |
2285 | * an in-core tree containing the contents of every gang header and the | |
2286 | * bps for every constituent of the gang block. | |
2287 | * | |
2288 | * With the gang tree now assembled, zio_gang_issue() just walks the gang tree | |
2289 | * and invokes a callback on each bp. To free a gang block, zio_gang_issue() | |
2290 | * calls zio_free_gang() -- a trivial wrapper around zio_free() -- for each bp. | |
2291 | * zio_claim_gang() provides a similarly trivial wrapper for zio_claim(). | |
2292 | * zio_read_gang() is a wrapper around zio_read() that omits reading gang | |
2293 | * headers, since we already have those in io_gang_tree. zio_rewrite_gang() | |
2294 | * performs a zio_rewrite() of the data or, for gang headers, a zio_rewrite() | |
2295 | * of the gang header plus zio_checksum_compute() of the data to update the | |
2296 | * gang header's blk_cksum as described above. | |
2297 | * | |
2298 | * The two-phase assemble/issue model solves the problem of partial failure -- | |
2299 | * what if you'd freed part of a gang block but then couldn't read the | |
2300 | * gang header for another part? Assembling the entire gang tree first | |
2301 | * ensures that all the necessary gang header I/O has succeeded before | |
2302 | * starting the actual work of free, claim, or write. Once the gang tree | |
2303 | * is assembled, free and claim are in-memory operations that cannot fail. | |
2304 | * | |
2305 | * In the event that a gang write fails, zio_dva_unallocate() walks the | |
2306 | * gang tree to immediately free (i.e. insert back into the space map) | |
2307 | * everything we've allocated. This ensures that we don't get ENOSPC | |
2308 | * errors during repeated suspend/resume cycles due to a flaky device. | |
2309 | * | |
2310 | * Gang rewrites only happen during sync-to-convergence. If we can't assemble | |
2311 | * the gang tree, we won't modify the block, so we can safely defer the free | |
2312 | * (knowing that the block is still intact). If we *can* assemble the gang | |
2313 | * tree, then even if some of the rewrites fail, zio_dva_unallocate() will free | |
2314 | * each constituent bp and we can allocate a new block on the next sync pass. | |
2315 | * | |
2316 | * In all cases, the gang tree allows complete recovery from partial failure. | |
34dc7c2f BB |
2317 | * ========================================================================== |
2318 | */ | |
b128c09f | 2319 | |
a6255b7f DQ |
2320 | static void |
2321 | zio_gang_issue_func_done(zio_t *zio) | |
2322 | { | |
2323 | abd_put(zio->io_abd); | |
2324 | } | |
2325 | ||
b128c09f | 2326 | static zio_t * |
a6255b7f DQ |
2327 | zio_read_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, |
2328 | uint64_t offset) | |
34dc7c2f | 2329 | { |
b128c09f BB |
2330 | if (gn != NULL) |
2331 | return (pio); | |
34dc7c2f | 2332 | |
a6255b7f DQ |
2333 | return (zio_read(pio, pio->io_spa, bp, abd_get_offset(data, offset), |
2334 | BP_GET_PSIZE(bp), zio_gang_issue_func_done, | |
2335 | NULL, pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
b128c09f BB |
2336 | &pio->io_bookmark)); |
2337 | } | |
2338 | ||
a6255b7f DQ |
2339 | static zio_t * |
2340 | zio_rewrite_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2341 | uint64_t offset) | |
b128c09f BB |
2342 | { |
2343 | zio_t *zio; | |
2344 | ||
2345 | if (gn != NULL) { | |
a6255b7f DQ |
2346 | abd_t *gbh_abd = |
2347 | abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
b128c09f | 2348 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2349 | gbh_abd, SPA_GANGBLOCKSIZE, zio_gang_issue_func_done, NULL, |
2350 | pio->io_priority, ZIO_GANG_CHILD_FLAGS(pio), | |
2351 | &pio->io_bookmark); | |
34dc7c2f | 2352 | /* |
b128c09f BB |
2353 | * As we rewrite each gang header, the pipeline will compute |
2354 | * a new gang block header checksum for it; but no one will | |
2355 | * compute a new data checksum, so we do that here. The one | |
2356 | * exception is the gang leader: the pipeline already computed | |
2357 | * its data checksum because that stage precedes gang assembly. | |
2358 | * (Presently, nothing actually uses interior data checksums; | |
2359 | * this is just good hygiene.) | |
34dc7c2f | 2360 | */ |
9babb374 | 2361 | if (gn != pio->io_gang_leader->io_gang_tree) { |
a6255b7f DQ |
2362 | abd_t *buf = abd_get_offset(data, offset); |
2363 | ||
b128c09f | 2364 | zio_checksum_compute(zio, BP_GET_CHECKSUM(bp), |
a6255b7f DQ |
2365 | buf, BP_GET_PSIZE(bp)); |
2366 | ||
2367 | abd_put(buf); | |
b128c09f | 2368 | } |
428870ff BB |
2369 | /* |
2370 | * If we are here to damage data for testing purposes, | |
2371 | * leave the GBH alone so that we can detect the damage. | |
2372 | */ | |
2373 | if (pio->io_gang_leader->io_flags & ZIO_FLAG_INDUCE_DAMAGE) | |
2374 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; | |
34dc7c2f | 2375 | } else { |
b128c09f | 2376 | zio = zio_rewrite(pio, pio->io_spa, pio->io_txg, bp, |
a6255b7f DQ |
2377 | abd_get_offset(data, offset), BP_GET_PSIZE(bp), |
2378 | zio_gang_issue_func_done, NULL, pio->io_priority, | |
b128c09f | 2379 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
34dc7c2f BB |
2380 | } |
2381 | ||
b128c09f BB |
2382 | return (zio); |
2383 | } | |
34dc7c2f | 2384 | |
b128c09f | 2385 | /* ARGSUSED */ |
a6255b7f DQ |
2386 | static zio_t * |
2387 | zio_free_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2388 | uint64_t offset) | |
b128c09f | 2389 | { |
428870ff BB |
2390 | return (zio_free_sync(pio, pio->io_spa, pio->io_txg, bp, |
2391 | ZIO_GANG_CHILD_FLAGS(pio))); | |
34dc7c2f BB |
2392 | } |
2393 | ||
b128c09f | 2394 | /* ARGSUSED */ |
a6255b7f DQ |
2395 | static zio_t * |
2396 | zio_claim_gang(zio_t *pio, blkptr_t *bp, zio_gang_node_t *gn, abd_t *data, | |
2397 | uint64_t offset) | |
34dc7c2f | 2398 | { |
b128c09f BB |
2399 | return (zio_claim(pio, pio->io_spa, pio->io_txg, bp, |
2400 | NULL, NULL, ZIO_GANG_CHILD_FLAGS(pio))); | |
2401 | } | |
2402 | ||
2403 | static zio_gang_issue_func_t *zio_gang_issue_func[ZIO_TYPES] = { | |
2404 | NULL, | |
2405 | zio_read_gang, | |
2406 | zio_rewrite_gang, | |
2407 | zio_free_gang, | |
2408 | zio_claim_gang, | |
2409 | NULL | |
2410 | }; | |
34dc7c2f | 2411 | |
b128c09f | 2412 | static void zio_gang_tree_assemble_done(zio_t *zio); |
34dc7c2f | 2413 | |
b128c09f BB |
2414 | static zio_gang_node_t * |
2415 | zio_gang_node_alloc(zio_gang_node_t **gnpp) | |
2416 | { | |
2417 | zio_gang_node_t *gn; | |
34dc7c2f | 2418 | |
b128c09f | 2419 | ASSERT(*gnpp == NULL); |
34dc7c2f | 2420 | |
79c76d5b | 2421 | gn = kmem_zalloc(sizeof (*gn), KM_SLEEP); |
b128c09f BB |
2422 | gn->gn_gbh = zio_buf_alloc(SPA_GANGBLOCKSIZE); |
2423 | *gnpp = gn; | |
34dc7c2f | 2424 | |
b128c09f | 2425 | return (gn); |
34dc7c2f BB |
2426 | } |
2427 | ||
34dc7c2f | 2428 | static void |
b128c09f | 2429 | zio_gang_node_free(zio_gang_node_t **gnpp) |
34dc7c2f | 2430 | { |
b128c09f | 2431 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2432 | |
1c27024e | 2433 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f BB |
2434 | ASSERT(gn->gn_child[g] == NULL); |
2435 | ||
2436 | zio_buf_free(gn->gn_gbh, SPA_GANGBLOCKSIZE); | |
2437 | kmem_free(gn, sizeof (*gn)); | |
2438 | *gnpp = NULL; | |
34dc7c2f BB |
2439 | } |
2440 | ||
b128c09f BB |
2441 | static void |
2442 | zio_gang_tree_free(zio_gang_node_t **gnpp) | |
34dc7c2f | 2443 | { |
b128c09f | 2444 | zio_gang_node_t *gn = *gnpp; |
34dc7c2f | 2445 | |
b128c09f BB |
2446 | if (gn == NULL) |
2447 | return; | |
34dc7c2f | 2448 | |
1c27024e | 2449 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) |
b128c09f | 2450 | zio_gang_tree_free(&gn->gn_child[g]); |
34dc7c2f | 2451 | |
b128c09f | 2452 | zio_gang_node_free(gnpp); |
34dc7c2f BB |
2453 | } |
2454 | ||
b128c09f | 2455 | static void |
9babb374 | 2456 | zio_gang_tree_assemble(zio_t *gio, blkptr_t *bp, zio_gang_node_t **gnpp) |
34dc7c2f | 2457 | { |
b128c09f | 2458 | zio_gang_node_t *gn = zio_gang_node_alloc(gnpp); |
a6255b7f | 2459 | abd_t *gbh_abd = abd_get_from_buf(gn->gn_gbh, SPA_GANGBLOCKSIZE); |
b128c09f | 2460 | |
9babb374 | 2461 | ASSERT(gio->io_gang_leader == gio); |
b128c09f | 2462 | ASSERT(BP_IS_GANG(bp)); |
34dc7c2f | 2463 | |
a6255b7f DQ |
2464 | zio_nowait(zio_read(gio, gio->io_spa, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2465 | zio_gang_tree_assemble_done, gn, gio->io_priority, | |
2466 | ZIO_GANG_CHILD_FLAGS(gio), &gio->io_bookmark)); | |
b128c09f | 2467 | } |
34dc7c2f | 2468 | |
b128c09f BB |
2469 | static void |
2470 | zio_gang_tree_assemble_done(zio_t *zio) | |
2471 | { | |
9babb374 | 2472 | zio_t *gio = zio->io_gang_leader; |
b128c09f BB |
2473 | zio_gang_node_t *gn = zio->io_private; |
2474 | blkptr_t *bp = zio->io_bp; | |
34dc7c2f | 2475 | |
9babb374 | 2476 | ASSERT(gio == zio_unique_parent(zio)); |
428870ff | 2477 | ASSERT(zio->io_child_count == 0); |
34dc7c2f | 2478 | |
b128c09f BB |
2479 | if (zio->io_error) |
2480 | return; | |
34dc7c2f | 2481 | |
a6255b7f | 2482 | /* this ABD was created from a linear buf in zio_gang_tree_assemble */ |
b128c09f | 2483 | if (BP_SHOULD_BYTESWAP(bp)) |
a6255b7f | 2484 | byteswap_uint64_array(abd_to_buf(zio->io_abd), zio->io_size); |
34dc7c2f | 2485 | |
a6255b7f | 2486 | ASSERT3P(abd_to_buf(zio->io_abd), ==, gn->gn_gbh); |
b128c09f | 2487 | ASSERT(zio->io_size == SPA_GANGBLOCKSIZE); |
428870ff | 2488 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2489 | |
a6255b7f DQ |
2490 | abd_put(zio->io_abd); |
2491 | ||
1c27024e | 2492 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2493 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2494 | if (!BP_IS_GANG(gbp)) | |
2495 | continue; | |
9babb374 | 2496 | zio_gang_tree_assemble(gio, gbp, &gn->gn_child[g]); |
b128c09f | 2497 | } |
34dc7c2f BB |
2498 | } |
2499 | ||
b128c09f | 2500 | static void |
a6255b7f DQ |
2501 | zio_gang_tree_issue(zio_t *pio, zio_gang_node_t *gn, blkptr_t *bp, abd_t *data, |
2502 | uint64_t offset) | |
34dc7c2f | 2503 | { |
9babb374 | 2504 | zio_t *gio = pio->io_gang_leader; |
b128c09f | 2505 | zio_t *zio; |
34dc7c2f | 2506 | |
b128c09f | 2507 | ASSERT(BP_IS_GANG(bp) == !!gn); |
9babb374 BB |
2508 | ASSERT(BP_GET_CHECKSUM(bp) == BP_GET_CHECKSUM(gio->io_bp)); |
2509 | ASSERT(BP_GET_LSIZE(bp) == BP_GET_PSIZE(bp) || gn == gio->io_gang_tree); | |
34dc7c2f | 2510 | |
b128c09f BB |
2511 | /* |
2512 | * If you're a gang header, your data is in gn->gn_gbh. | |
2513 | * If you're a gang member, your data is in 'data' and gn == NULL. | |
2514 | */ | |
a6255b7f | 2515 | zio = zio_gang_issue_func[gio->io_type](pio, bp, gn, data, offset); |
34dc7c2f | 2516 | |
b128c09f | 2517 | if (gn != NULL) { |
428870ff | 2518 | ASSERT(gn->gn_gbh->zg_tail.zec_magic == ZEC_MAGIC); |
34dc7c2f | 2519 | |
1c27024e | 2520 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
2521 | blkptr_t *gbp = &gn->gn_gbh->zg_blkptr[g]; |
2522 | if (BP_IS_HOLE(gbp)) | |
2523 | continue; | |
a6255b7f DQ |
2524 | zio_gang_tree_issue(zio, gn->gn_child[g], gbp, data, |
2525 | offset); | |
2526 | offset += BP_GET_PSIZE(gbp); | |
b128c09f | 2527 | } |
34dc7c2f BB |
2528 | } |
2529 | ||
9babb374 | 2530 | if (gn == gio->io_gang_tree) |
a6255b7f | 2531 | ASSERT3U(gio->io_size, ==, offset); |
34dc7c2f | 2532 | |
b128c09f BB |
2533 | if (zio != pio) |
2534 | zio_nowait(zio); | |
34dc7c2f BB |
2535 | } |
2536 | ||
62840030 | 2537 | static zio_t * |
b128c09f | 2538 | zio_gang_assemble(zio_t *zio) |
34dc7c2f | 2539 | { |
b128c09f | 2540 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2541 | |
9babb374 BB |
2542 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == NULL); |
2543 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
2544 | ||
2545 | zio->io_gang_leader = zio; | |
34dc7c2f | 2546 | |
b128c09f | 2547 | zio_gang_tree_assemble(zio, bp, &zio->io_gang_tree); |
34dc7c2f | 2548 | |
62840030 | 2549 | return (zio); |
34dc7c2f BB |
2550 | } |
2551 | ||
62840030 | 2552 | static zio_t * |
b128c09f | 2553 | zio_gang_issue(zio_t *zio) |
34dc7c2f | 2554 | { |
b128c09f | 2555 | blkptr_t *bp = zio->io_bp; |
34dc7c2f | 2556 | |
ddc751d5 | 2557 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2558 | return (NULL); |
ddc751d5 | 2559 | } |
34dc7c2f | 2560 | |
9babb374 BB |
2561 | ASSERT(BP_IS_GANG(bp) && zio->io_gang_leader == zio); |
2562 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
34dc7c2f | 2563 | |
b128c09f | 2564 | if (zio->io_child_error[ZIO_CHILD_GANG] == 0) |
a6255b7f DQ |
2565 | zio_gang_tree_issue(zio, zio->io_gang_tree, bp, zio->io_abd, |
2566 | 0); | |
b128c09f | 2567 | else |
9babb374 | 2568 | zio_gang_tree_free(&zio->io_gang_tree); |
34dc7c2f | 2569 | |
b128c09f | 2570 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 2571 | |
62840030 | 2572 | return (zio); |
34dc7c2f BB |
2573 | } |
2574 | ||
2575 | static void | |
b128c09f | 2576 | zio_write_gang_member_ready(zio_t *zio) |
34dc7c2f | 2577 | { |
d164b209 | 2578 | zio_t *pio = zio_unique_parent(zio); |
34dc7c2f BB |
2579 | dva_t *cdva = zio->io_bp->blk_dva; |
2580 | dva_t *pdva = pio->io_bp->blk_dva; | |
2581 | uint64_t asize; | |
d1d7e268 | 2582 | ASSERTV(zio_t *gio = zio->io_gang_leader); |
34dc7c2f | 2583 | |
b128c09f BB |
2584 | if (BP_IS_HOLE(zio->io_bp)) |
2585 | return; | |
2586 | ||
2587 | ASSERT(BP_IS_HOLE(&zio->io_bp_orig)); | |
2588 | ||
2589 | ASSERT(zio->io_child_type == ZIO_CHILD_GANG); | |
428870ff BB |
2590 | ASSERT3U(zio->io_prop.zp_copies, ==, gio->io_prop.zp_copies); |
2591 | ASSERT3U(zio->io_prop.zp_copies, <=, BP_GET_NDVAS(zio->io_bp)); | |
2592 | ASSERT3U(pio->io_prop.zp_copies, <=, BP_GET_NDVAS(pio->io_bp)); | |
34dc7c2f | 2593 | ASSERT3U(BP_GET_NDVAS(zio->io_bp), <=, BP_GET_NDVAS(pio->io_bp)); |
34dc7c2f BB |
2594 | |
2595 | mutex_enter(&pio->io_lock); | |
1c27024e | 2596 | for (int d = 0; d < BP_GET_NDVAS(zio->io_bp); d++) { |
34dc7c2f BB |
2597 | ASSERT(DVA_GET_GANG(&pdva[d])); |
2598 | asize = DVA_GET_ASIZE(&pdva[d]); | |
2599 | asize += DVA_GET_ASIZE(&cdva[d]); | |
2600 | DVA_SET_ASIZE(&pdva[d], asize); | |
2601 | } | |
2602 | mutex_exit(&pio->io_lock); | |
2603 | } | |
2604 | ||
a6255b7f DQ |
2605 | static void |
2606 | zio_write_gang_done(zio_t *zio) | |
2607 | { | |
c955398b BL |
2608 | /* |
2609 | * The io_abd field will be NULL for a zio with no data. The io_flags | |
2610 | * will initially have the ZIO_FLAG_NODATA bit flag set, but we can't | |
2611 | * check for it here as it is cleared in zio_ready. | |
2612 | */ | |
2613 | if (zio->io_abd != NULL) | |
2614 | abd_put(zio->io_abd); | |
a6255b7f DQ |
2615 | } |
2616 | ||
62840030 | 2617 | static zio_t * |
b128c09f | 2618 | zio_write_gang_block(zio_t *pio) |
34dc7c2f | 2619 | { |
b128c09f | 2620 | spa_t *spa = pio->io_spa; |
3dfb57a3 | 2621 | metaslab_class_t *mc = spa_normal_class(spa); |
b128c09f | 2622 | blkptr_t *bp = pio->io_bp; |
9babb374 | 2623 | zio_t *gio = pio->io_gang_leader; |
b128c09f BB |
2624 | zio_t *zio; |
2625 | zio_gang_node_t *gn, **gnpp; | |
34dc7c2f | 2626 | zio_gbh_phys_t *gbh; |
a6255b7f | 2627 | abd_t *gbh_abd; |
b128c09f BB |
2628 | uint64_t txg = pio->io_txg; |
2629 | uint64_t resid = pio->io_size; | |
2630 | uint64_t lsize; | |
428870ff | 2631 | int copies = gio->io_prop.zp_copies; |
b5256303 | 2632 | int gbh_copies; |
b128c09f | 2633 | zio_prop_t zp; |
1c27024e | 2634 | int error; |
c955398b | 2635 | boolean_t has_data = !(pio->io_flags & ZIO_FLAG_NODATA); |
b5256303 TC |
2636 | |
2637 | /* | |
2638 | * encrypted blocks need DVA[2] free so encrypted gang headers can't | |
2639 | * have a third copy. | |
2640 | */ | |
2641 | gbh_copies = MIN(copies + 1, spa_max_replication(spa)); | |
2642 | if (gio->io_prop.zp_encrypt && gbh_copies >= SPA_DVAS_PER_BP) | |
2643 | gbh_copies = SPA_DVAS_PER_BP - 1; | |
2644 | ||
1c27024e | 2645 | int flags = METASLAB_HINTBP_FAVOR | METASLAB_GANG_HEADER; |
3dfb57a3 DB |
2646 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2647 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2648 | ASSERT(has_data); |
3dfb57a3 DB |
2649 | |
2650 | flags |= METASLAB_ASYNC_ALLOC; | |
424fd7c3 | 2651 | VERIFY(zfs_refcount_held(&mc->mc_alloc_slots[pio->io_allocator], |
492f64e9 | 2652 | pio)); |
3dfb57a3 DB |
2653 | |
2654 | /* | |
2655 | * The logical zio has already placed a reservation for | |
2656 | * 'copies' allocation slots but gang blocks may require | |
2657 | * additional copies. These additional copies | |
2658 | * (i.e. gbh_copies - copies) are guaranteed to succeed | |
2659 | * since metaslab_class_throttle_reserve() always allows | |
2660 | * additional reservations for gang blocks. | |
2661 | */ | |
2662 | VERIFY(metaslab_class_throttle_reserve(mc, gbh_copies - copies, | |
492f64e9 | 2663 | pio->io_allocator, pio, flags)); |
3dfb57a3 DB |
2664 | } |
2665 | ||
2666 | error = metaslab_alloc(spa, mc, SPA_GANGBLOCKSIZE, | |
4e21fd06 | 2667 | bp, gbh_copies, txg, pio == gio ? NULL : gio->io_bp, flags, |
492f64e9 | 2668 | &pio->io_alloc_list, pio, pio->io_allocator); |
34dc7c2f | 2669 | if (error) { |
3dfb57a3 DB |
2670 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
2671 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2672 | ASSERT(has_data); |
3dfb57a3 DB |
2673 | |
2674 | /* | |
2675 | * If we failed to allocate the gang block header then | |
2676 | * we remove any additional allocation reservations that | |
2677 | * we placed here. The original reservation will | |
2678 | * be removed when the logical I/O goes to the ready | |
2679 | * stage. | |
2680 | */ | |
2681 | metaslab_class_throttle_unreserve(mc, | |
492f64e9 | 2682 | gbh_copies - copies, pio->io_allocator, pio); |
3dfb57a3 DB |
2683 | } |
2684 | ||
b128c09f | 2685 | pio->io_error = error; |
62840030 | 2686 | return (pio); |
34dc7c2f BB |
2687 | } |
2688 | ||
9babb374 BB |
2689 | if (pio == gio) { |
2690 | gnpp = &gio->io_gang_tree; | |
b128c09f BB |
2691 | } else { |
2692 | gnpp = pio->io_private; | |
2693 | ASSERT(pio->io_ready == zio_write_gang_member_ready); | |
34dc7c2f BB |
2694 | } |
2695 | ||
b128c09f BB |
2696 | gn = zio_gang_node_alloc(gnpp); |
2697 | gbh = gn->gn_gbh; | |
2698 | bzero(gbh, SPA_GANGBLOCKSIZE); | |
a6255b7f | 2699 | gbh_abd = abd_get_from_buf(gbh, SPA_GANGBLOCKSIZE); |
34dc7c2f | 2700 | |
b128c09f BB |
2701 | /* |
2702 | * Create the gang header. | |
2703 | */ | |
a6255b7f DQ |
2704 | zio = zio_rewrite(pio, spa, txg, bp, gbh_abd, SPA_GANGBLOCKSIZE, |
2705 | zio_write_gang_done, NULL, pio->io_priority, | |
2706 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); | |
34dc7c2f | 2707 | |
b128c09f BB |
2708 | /* |
2709 | * Create and nowait the gang children. | |
2710 | */ | |
1c27024e | 2711 | for (int g = 0; resid != 0; resid -= lsize, g++) { |
b128c09f BB |
2712 | lsize = P2ROUNDUP(resid / (SPA_GBH_NBLKPTRS - g), |
2713 | SPA_MINBLOCKSIZE); | |
2714 | ASSERT(lsize >= SPA_MINBLOCKSIZE && lsize <= resid); | |
2715 | ||
9babb374 | 2716 | zp.zp_checksum = gio->io_prop.zp_checksum; |
b128c09f BB |
2717 | zp.zp_compress = ZIO_COMPRESS_OFF; |
2718 | zp.zp_type = DMU_OT_NONE; | |
2719 | zp.zp_level = 0; | |
428870ff | 2720 | zp.zp_copies = gio->io_prop.zp_copies; |
03c6040b GW |
2721 | zp.zp_dedup = B_FALSE; |
2722 | zp.zp_dedup_verify = B_FALSE; | |
2723 | zp.zp_nopwrite = B_FALSE; | |
4807c0ba TC |
2724 | zp.zp_encrypt = gio->io_prop.zp_encrypt; |
2725 | zp.zp_byteorder = gio->io_prop.zp_byteorder; | |
b5256303 TC |
2726 | bzero(zp.zp_salt, ZIO_DATA_SALT_LEN); |
2727 | bzero(zp.zp_iv, ZIO_DATA_IV_LEN); | |
2728 | bzero(zp.zp_mac, ZIO_DATA_MAC_LEN); | |
b128c09f | 2729 | |
1c27024e | 2730 | zio_t *cio = zio_write(zio, spa, txg, &gbh->zg_blkptr[g], |
c955398b BL |
2731 | has_data ? abd_get_offset(pio->io_abd, pio->io_size - |
2732 | resid) : NULL, lsize, lsize, &zp, | |
2733 | zio_write_gang_member_ready, NULL, NULL, | |
a6255b7f | 2734 | zio_write_gang_done, &gn->gn_child[g], pio->io_priority, |
3dfb57a3 DB |
2735 | ZIO_GANG_CHILD_FLAGS(pio), &pio->io_bookmark); |
2736 | ||
2737 | if (pio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
2738 | ASSERT(pio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
c955398b | 2739 | ASSERT(has_data); |
3dfb57a3 DB |
2740 | |
2741 | /* | |
2742 | * Gang children won't throttle but we should | |
2743 | * account for their work, so reserve an allocation | |
2744 | * slot for them here. | |
2745 | */ | |
2746 | VERIFY(metaslab_class_throttle_reserve(mc, | |
492f64e9 | 2747 | zp.zp_copies, cio->io_allocator, cio, flags)); |
3dfb57a3 DB |
2748 | } |
2749 | zio_nowait(cio); | |
b128c09f | 2750 | } |
34dc7c2f BB |
2751 | |
2752 | /* | |
b128c09f | 2753 | * Set pio's pipeline to just wait for zio to finish. |
34dc7c2f | 2754 | */ |
b128c09f BB |
2755 | pio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
2756 | ||
920dd524 ED |
2757 | /* |
2758 | * We didn't allocate this bp, so make sure it doesn't get unmarked. | |
2759 | */ | |
2760 | pio->io_flags &= ~ZIO_FLAG_FASTWRITE; | |
2761 | ||
b128c09f BB |
2762 | zio_nowait(zio); |
2763 | ||
62840030 | 2764 | return (pio); |
34dc7c2f BB |
2765 | } |
2766 | ||
03c6040b | 2767 | /* |
3c67d83a TH |
2768 | * The zio_nop_write stage in the pipeline determines if allocating a |
2769 | * new bp is necessary. The nopwrite feature can handle writes in | |
2770 | * either syncing or open context (i.e. zil writes) and as a result is | |
2771 | * mutually exclusive with dedup. | |
2772 | * | |
2773 | * By leveraging a cryptographically secure checksum, such as SHA256, we | |
2774 | * can compare the checksums of the new data and the old to determine if | |
2775 | * allocating a new block is required. Note that our requirements for | |
2776 | * cryptographic strength are fairly weak: there can't be any accidental | |
2777 | * hash collisions, but we don't need to be secure against intentional | |
2778 | * (malicious) collisions. To trigger a nopwrite, you have to be able | |
2779 | * to write the file to begin with, and triggering an incorrect (hash | |
2780 | * collision) nopwrite is no worse than simply writing to the file. | |
2781 | * That said, there are no known attacks against the checksum algorithms | |
2782 | * used for nopwrite, assuming that the salt and the checksums | |
2783 | * themselves remain secret. | |
03c6040b | 2784 | */ |
62840030 | 2785 | static zio_t * |
03c6040b GW |
2786 | zio_nop_write(zio_t *zio) |
2787 | { | |
2788 | blkptr_t *bp = zio->io_bp; | |
2789 | blkptr_t *bp_orig = &zio->io_bp_orig; | |
2790 | zio_prop_t *zp = &zio->io_prop; | |
2791 | ||
2792 | ASSERT(BP_GET_LEVEL(bp) == 0); | |
2793 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
2794 | ASSERT(zp->zp_nopwrite); | |
2795 | ASSERT(!zp->zp_dedup); | |
2796 | ASSERT(zio->io_bp_override == NULL); | |
2797 | ASSERT(IO_IS_ALLOCATING(zio)); | |
2798 | ||
2799 | /* | |
2800 | * Check to see if the original bp and the new bp have matching | |
2801 | * characteristics (i.e. same checksum, compression algorithms, etc). | |
2802 | * If they don't then just continue with the pipeline which will | |
2803 | * allocate a new bp. | |
2804 | */ | |
2805 | if (BP_IS_HOLE(bp_orig) || | |
3c67d83a TH |
2806 | !(zio_checksum_table[BP_GET_CHECKSUM(bp)].ci_flags & |
2807 | ZCHECKSUM_FLAG_NOPWRITE) || | |
b5256303 | 2808 | BP_IS_ENCRYPTED(bp) || BP_IS_ENCRYPTED(bp_orig) || |
03c6040b GW |
2809 | BP_GET_CHECKSUM(bp) != BP_GET_CHECKSUM(bp_orig) || |
2810 | BP_GET_COMPRESS(bp) != BP_GET_COMPRESS(bp_orig) || | |
2811 | BP_GET_DEDUP(bp) != BP_GET_DEDUP(bp_orig) || | |
2812 | zp->zp_copies != BP_GET_NDVAS(bp_orig)) | |
62840030 | 2813 | return (zio); |
03c6040b GW |
2814 | |
2815 | /* | |
2816 | * If the checksums match then reset the pipeline so that we | |
2817 | * avoid allocating a new bp and issuing any I/O. | |
2818 | */ | |
2819 | if (ZIO_CHECKSUM_EQUAL(bp->blk_cksum, bp_orig->blk_cksum)) { | |
3c67d83a TH |
2820 | ASSERT(zio_checksum_table[zp->zp_checksum].ci_flags & |
2821 | ZCHECKSUM_FLAG_NOPWRITE); | |
03c6040b GW |
2822 | ASSERT3U(BP_GET_PSIZE(bp), ==, BP_GET_PSIZE(bp_orig)); |
2823 | ASSERT3U(BP_GET_LSIZE(bp), ==, BP_GET_LSIZE(bp_orig)); | |
2824 | ASSERT(zp->zp_compress != ZIO_COMPRESS_OFF); | |
2825 | ASSERT(bcmp(&bp->blk_prop, &bp_orig->blk_prop, | |
2826 | sizeof (uint64_t)) == 0); | |
2827 | ||
2828 | *bp = *bp_orig; | |
2829 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
2830 | zio->io_flags |= ZIO_FLAG_NOPWRITE; | |
2831 | } | |
2832 | ||
62840030 | 2833 | return (zio); |
03c6040b GW |
2834 | } |
2835 | ||
34dc7c2f BB |
2836 | /* |
2837 | * ========================================================================== | |
428870ff | 2838 | * Dedup |
34dc7c2f BB |
2839 | * ========================================================================== |
2840 | */ | |
428870ff BB |
2841 | static void |
2842 | zio_ddt_child_read_done(zio_t *zio) | |
2843 | { | |
2844 | blkptr_t *bp = zio->io_bp; | |
2845 | ddt_entry_t *dde = zio->io_private; | |
2846 | ddt_phys_t *ddp; | |
2847 | zio_t *pio = zio_unique_parent(zio); | |
2848 | ||
2849 | mutex_enter(&pio->io_lock); | |
2850 | ddp = ddt_phys_select(dde, bp); | |
2851 | if (zio->io_error == 0) | |
2852 | ddt_phys_clear(ddp); /* this ddp doesn't need repair */ | |
a6255b7f DQ |
2853 | |
2854 | if (zio->io_error == 0 && dde->dde_repair_abd == NULL) | |
2855 | dde->dde_repair_abd = zio->io_abd; | |
428870ff | 2856 | else |
a6255b7f | 2857 | abd_free(zio->io_abd); |
428870ff BB |
2858 | mutex_exit(&pio->io_lock); |
2859 | } | |
2860 | ||
62840030 | 2861 | static zio_t * |
428870ff BB |
2862 | zio_ddt_read_start(zio_t *zio) |
2863 | { | |
2864 | blkptr_t *bp = zio->io_bp; | |
2865 | ||
2866 | ASSERT(BP_GET_DEDUP(bp)); | |
2867 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2868 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2869 | ||
2870 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2871 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2872 | ddt_entry_t *dde = ddt_repair_start(ddt, bp); | |
2873 | ddt_phys_t *ddp = dde->dde_phys; | |
2874 | ddt_phys_t *ddp_self = ddt_phys_select(dde, bp); | |
2875 | blkptr_t blk; | |
2876 | ||
2877 | ASSERT(zio->io_vsd == NULL); | |
2878 | zio->io_vsd = dde; | |
2879 | ||
2880 | if (ddp_self == NULL) | |
62840030 | 2881 | return (zio); |
428870ff | 2882 | |
1c27024e | 2883 | for (int p = 0; p < DDT_PHYS_TYPES; p++, ddp++) { |
428870ff BB |
2884 | if (ddp->ddp_phys_birth == 0 || ddp == ddp_self) |
2885 | continue; | |
2886 | ddt_bp_create(ddt->ddt_checksum, &dde->dde_key, ddp, | |
2887 | &blk); | |
2888 | zio_nowait(zio_read(zio, zio->io_spa, &blk, | |
a6255b7f DQ |
2889 | abd_alloc_for_io(zio->io_size, B_TRUE), |
2890 | zio->io_size, zio_ddt_child_read_done, dde, | |
2891 | zio->io_priority, ZIO_DDT_CHILD_FLAGS(zio) | | |
2892 | ZIO_FLAG_DONT_PROPAGATE, &zio->io_bookmark)); | |
428870ff | 2893 | } |
62840030 | 2894 | return (zio); |
428870ff BB |
2895 | } |
2896 | ||
2897 | zio_nowait(zio_read(zio, zio->io_spa, bp, | |
a6255b7f | 2898 | zio->io_abd, zio->io_size, NULL, NULL, zio->io_priority, |
428870ff BB |
2899 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark)); |
2900 | ||
62840030 | 2901 | return (zio); |
428870ff BB |
2902 | } |
2903 | ||
62840030 | 2904 | static zio_t * |
428870ff BB |
2905 | zio_ddt_read_done(zio_t *zio) |
2906 | { | |
2907 | blkptr_t *bp = zio->io_bp; | |
2908 | ||
ddc751d5 | 2909 | if (zio_wait_for_children(zio, ZIO_CHILD_DDT_BIT, ZIO_WAIT_DONE)) { |
62840030 | 2910 | return (NULL); |
ddc751d5 | 2911 | } |
428870ff BB |
2912 | |
2913 | ASSERT(BP_GET_DEDUP(bp)); | |
2914 | ASSERT(BP_GET_PSIZE(bp) == zio->io_size); | |
2915 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
2916 | ||
2917 | if (zio->io_child_error[ZIO_CHILD_DDT]) { | |
2918 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
2919 | ddt_entry_t *dde = zio->io_vsd; | |
2920 | if (ddt == NULL) { | |
2921 | ASSERT(spa_load_state(zio->io_spa) != SPA_LOAD_NONE); | |
62840030 | 2922 | return (zio); |
428870ff BB |
2923 | } |
2924 | if (dde == NULL) { | |
2925 | zio->io_stage = ZIO_STAGE_DDT_READ_START >> 1; | |
2926 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_FALSE); | |
62840030 | 2927 | return (NULL); |
428870ff | 2928 | } |
a6255b7f DQ |
2929 | if (dde->dde_repair_abd != NULL) { |
2930 | abd_copy(zio->io_abd, dde->dde_repair_abd, | |
2931 | zio->io_size); | |
428870ff BB |
2932 | zio->io_child_error[ZIO_CHILD_DDT] = 0; |
2933 | } | |
2934 | ddt_repair_done(ddt, dde); | |
2935 | zio->io_vsd = NULL; | |
2936 | } | |
2937 | ||
2938 | ASSERT(zio->io_vsd == NULL); | |
2939 | ||
62840030 | 2940 | return (zio); |
428870ff BB |
2941 | } |
2942 | ||
2943 | static boolean_t | |
2944 | zio_ddt_collision(zio_t *zio, ddt_t *ddt, ddt_entry_t *dde) | |
2945 | { | |
2946 | spa_t *spa = zio->io_spa; | |
c17bcf83 | 2947 | boolean_t do_raw = !!(zio->io_flags & ZIO_FLAG_RAW); |
428870ff | 2948 | |
c17bcf83 | 2949 | ASSERT(!(zio->io_bp_override && do_raw)); |
2aa34383 | 2950 | |
428870ff BB |
2951 | /* |
2952 | * Note: we compare the original data, not the transformed data, | |
2953 | * because when zio->io_bp is an override bp, we will not have | |
2954 | * pushed the I/O transforms. That's an important optimization | |
2955 | * because otherwise we'd compress/encrypt all dmu_sync() data twice. | |
c17bcf83 | 2956 | * However, we should never get a raw, override zio so in these |
b5256303 | 2957 | * cases we can compare the io_abd directly. This is useful because |
c17bcf83 TC |
2958 | * it allows us to do dedup verification even if we don't have access |
2959 | * to the original data (for instance, if the encryption keys aren't | |
2960 | * loaded). | |
428870ff | 2961 | */ |
c17bcf83 | 2962 | |
1c27024e | 2963 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2964 | zio_t *lio = dde->dde_lead_zio[p]; |
2965 | ||
c17bcf83 TC |
2966 | if (lio != NULL && do_raw) { |
2967 | return (lio->io_size != zio->io_size || | |
a6255b7f | 2968 | abd_cmp(zio->io_abd, lio->io_abd) != 0); |
c17bcf83 | 2969 | } else if (lio != NULL) { |
428870ff | 2970 | return (lio->io_orig_size != zio->io_orig_size || |
a6255b7f | 2971 | abd_cmp(zio->io_orig_abd, lio->io_orig_abd) != 0); |
428870ff BB |
2972 | } |
2973 | } | |
2974 | ||
1c27024e | 2975 | for (int p = DDT_PHYS_SINGLE; p <= DDT_PHYS_TRIPLE; p++) { |
428870ff BB |
2976 | ddt_phys_t *ddp = &dde->dde_phys[p]; |
2977 | ||
c17bcf83 TC |
2978 | if (ddp->ddp_phys_birth != 0 && do_raw) { |
2979 | blkptr_t blk = *zio->io_bp; | |
2980 | uint64_t psize; | |
a6255b7f | 2981 | abd_t *tmpabd; |
c17bcf83 TC |
2982 | int error; |
2983 | ||
2984 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
2985 | psize = BP_GET_PSIZE(&blk); | |
2986 | ||
2987 | if (psize != zio->io_size) | |
2988 | return (B_TRUE); | |
2989 | ||
2990 | ddt_exit(ddt); | |
2991 | ||
a6255b7f | 2992 | tmpabd = abd_alloc_for_io(psize, B_TRUE); |
c17bcf83 | 2993 | |
a6255b7f | 2994 | error = zio_wait(zio_read(NULL, spa, &blk, tmpabd, |
c17bcf83 TC |
2995 | psize, NULL, NULL, ZIO_PRIORITY_SYNC_READ, |
2996 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | | |
2997 | ZIO_FLAG_RAW, &zio->io_bookmark)); | |
2998 | ||
2999 | if (error == 0) { | |
a6255b7f | 3000 | if (abd_cmp(tmpabd, zio->io_abd) != 0) |
c17bcf83 TC |
3001 | error = SET_ERROR(ENOENT); |
3002 | } | |
3003 | ||
a6255b7f | 3004 | abd_free(tmpabd); |
c17bcf83 TC |
3005 | ddt_enter(ddt); |
3006 | return (error != 0); | |
3007 | } else if (ddp->ddp_phys_birth != 0) { | |
428870ff | 3008 | arc_buf_t *abuf = NULL; |
2a432414 | 3009 | arc_flags_t aflags = ARC_FLAG_WAIT; |
428870ff BB |
3010 | blkptr_t blk = *zio->io_bp; |
3011 | int error; | |
3012 | ||
3013 | ddt_bp_fill(ddp, &blk, ddp->ddp_phys_birth); | |
3014 | ||
c17bcf83 TC |
3015 | if (BP_GET_LSIZE(&blk) != zio->io_orig_size) |
3016 | return (B_TRUE); | |
3017 | ||
428870ff BB |
3018 | ddt_exit(ddt); |
3019 | ||
294f6806 | 3020 | error = arc_read(NULL, spa, &blk, |
428870ff BB |
3021 | arc_getbuf_func, &abuf, ZIO_PRIORITY_SYNC_READ, |
3022 | ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE, | |
3023 | &aflags, &zio->io_bookmark); | |
3024 | ||
3025 | if (error == 0) { | |
a6255b7f | 3026 | if (abd_cmp_buf(zio->io_orig_abd, abuf->b_data, |
428870ff | 3027 | zio->io_orig_size) != 0) |
c17bcf83 | 3028 | error = SET_ERROR(ENOENT); |
d3c2ae1c | 3029 | arc_buf_destroy(abuf, &abuf); |
428870ff BB |
3030 | } |
3031 | ||
3032 | ddt_enter(ddt); | |
3033 | return (error != 0); | |
3034 | } | |
3035 | } | |
3036 | ||
3037 | return (B_FALSE); | |
3038 | } | |
3039 | ||
3040 | static void | |
3041 | zio_ddt_child_write_ready(zio_t *zio) | |
3042 | { | |
3043 | int p = zio->io_prop.zp_copies; | |
3044 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3045 | ddt_entry_t *dde = zio->io_private; | |
3046 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3047 | zio_t *pio; | |
3048 | ||
3049 | if (zio->io_error) | |
3050 | return; | |
3051 | ||
3052 | ddt_enter(ddt); | |
3053 | ||
3054 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3055 | ||
3056 | ddt_phys_fill(ddp, zio->io_bp); | |
3057 | ||
1c27024e | 3058 | zio_link_t *zl = NULL; |
3dfb57a3 | 3059 | while ((pio = zio_walk_parents(zio, &zl)) != NULL) |
428870ff BB |
3060 | ddt_bp_fill(ddp, pio->io_bp, zio->io_txg); |
3061 | ||
3062 | ddt_exit(ddt); | |
3063 | } | |
3064 | ||
3065 | static void | |
3066 | zio_ddt_child_write_done(zio_t *zio) | |
3067 | { | |
3068 | int p = zio->io_prop.zp_copies; | |
3069 | ddt_t *ddt = ddt_select(zio->io_spa, zio->io_bp); | |
3070 | ddt_entry_t *dde = zio->io_private; | |
3071 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3072 | ||
3073 | ddt_enter(ddt); | |
3074 | ||
3075 | ASSERT(ddp->ddp_refcnt == 0); | |
3076 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3077 | dde->dde_lead_zio[p] = NULL; | |
3078 | ||
3079 | if (zio->io_error == 0) { | |
3dfb57a3 DB |
3080 | zio_link_t *zl = NULL; |
3081 | while (zio_walk_parents(zio, &zl) != NULL) | |
428870ff BB |
3082 | ddt_phys_addref(ddp); |
3083 | } else { | |
3084 | ddt_phys_clear(ddp); | |
3085 | } | |
3086 | ||
3087 | ddt_exit(ddt); | |
3088 | } | |
3089 | ||
3090 | static void | |
3091 | zio_ddt_ditto_write_done(zio_t *zio) | |
3092 | { | |
3093 | int p = DDT_PHYS_DITTO; | |
1c27024e | 3094 | ASSERTV(zio_prop_t *zp = &zio->io_prop); |
428870ff BB |
3095 | blkptr_t *bp = zio->io_bp; |
3096 | ddt_t *ddt = ddt_select(zio->io_spa, bp); | |
3097 | ddt_entry_t *dde = zio->io_private; | |
3098 | ddt_phys_t *ddp = &dde->dde_phys[p]; | |
3099 | ddt_key_t *ddk = &dde->dde_key; | |
3100 | ||
3101 | ddt_enter(ddt); | |
3102 | ||
3103 | ASSERT(ddp->ddp_refcnt == 0); | |
3104 | ASSERT(dde->dde_lead_zio[p] == zio); | |
3105 | dde->dde_lead_zio[p] = NULL; | |
3106 | ||
3107 | if (zio->io_error == 0) { | |
3108 | ASSERT(ZIO_CHECKSUM_EQUAL(bp->blk_cksum, ddk->ddk_cksum)); | |
3109 | ASSERT(zp->zp_copies < SPA_DVAS_PER_BP); | |
3110 | ASSERT(zp->zp_copies == BP_GET_NDVAS(bp) - BP_IS_GANG(bp)); | |
3111 | if (ddp->ddp_phys_birth != 0) | |
3112 | ddt_phys_free(ddt, ddk, ddp, zio->io_txg); | |
3113 | ddt_phys_fill(ddp, bp); | |
3114 | } | |
3115 | ||
3116 | ddt_exit(ddt); | |
3117 | } | |
3118 | ||
62840030 | 3119 | static zio_t * |
428870ff BB |
3120 | zio_ddt_write(zio_t *zio) |
3121 | { | |
3122 | spa_t *spa = zio->io_spa; | |
3123 | blkptr_t *bp = zio->io_bp; | |
3124 | uint64_t txg = zio->io_txg; | |
3125 | zio_prop_t *zp = &zio->io_prop; | |
3126 | int p = zp->zp_copies; | |
3127 | int ditto_copies; | |
3128 | zio_t *cio = NULL; | |
3129 | zio_t *dio = NULL; | |
3130 | ddt_t *ddt = ddt_select(spa, bp); | |
3131 | ddt_entry_t *dde; | |
3132 | ddt_phys_t *ddp; | |
3133 | ||
3134 | ASSERT(BP_GET_DEDUP(bp)); | |
3135 | ASSERT(BP_GET_CHECKSUM(bp) == zp->zp_checksum); | |
3136 | ASSERT(BP_IS_HOLE(bp) || zio->io_bp_override); | |
c17bcf83 | 3137 | ASSERT(!(zio->io_bp_override && (zio->io_flags & ZIO_FLAG_RAW))); |
428870ff BB |
3138 | |
3139 | ddt_enter(ddt); | |
3140 | dde = ddt_lookup(ddt, bp, B_TRUE); | |
3141 | ddp = &dde->dde_phys[p]; | |
3142 | ||
3143 | if (zp->zp_dedup_verify && zio_ddt_collision(zio, ddt, dde)) { | |
3144 | /* | |
3145 | * If we're using a weak checksum, upgrade to a strong checksum | |
3146 | * and try again. If we're already using a strong checksum, | |
3147 | * we can't resolve it, so just convert to an ordinary write. | |
3148 | * (And automatically e-mail a paper to Nature?) | |
3149 | */ | |
3c67d83a TH |
3150 | if (!(zio_checksum_table[zp->zp_checksum].ci_flags & |
3151 | ZCHECKSUM_FLAG_DEDUP)) { | |
428870ff BB |
3152 | zp->zp_checksum = spa_dedup_checksum(spa); |
3153 | zio_pop_transforms(zio); | |
3154 | zio->io_stage = ZIO_STAGE_OPEN; | |
3155 | BP_ZERO(bp); | |
3156 | } else { | |
03c6040b | 3157 | zp->zp_dedup = B_FALSE; |
428870ff BB |
3158 | } |
3159 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
3160 | ddt_exit(ddt); | |
62840030 | 3161 | return (zio); |
428870ff BB |
3162 | } |
3163 | ||
3164 | ditto_copies = ddt_ditto_copies_needed(ddt, dde, ddp); | |
3165 | ASSERT(ditto_copies < SPA_DVAS_PER_BP); | |
3166 | ||
3167 | if (ditto_copies > ddt_ditto_copies_present(dde) && | |
3168 | dde->dde_lead_zio[DDT_PHYS_DITTO] == NULL) { | |
3169 | zio_prop_t czp = *zp; | |
3170 | ||
3171 | czp.zp_copies = ditto_copies; | |
3172 | ||
3173 | /* | |
3174 | * If we arrived here with an override bp, we won't have run | |
3175 | * the transform stack, so we won't have the data we need to | |
3176 | * generate a child i/o. So, toss the override bp and restart. | |
3177 | * This is safe, because using the override bp is just an | |
3178 | * optimization; and it's rare, so the cost doesn't matter. | |
3179 | */ | |
3180 | if (zio->io_bp_override) { | |
3181 | zio_pop_transforms(zio); | |
3182 | zio->io_stage = ZIO_STAGE_OPEN; | |
3183 | zio->io_pipeline = ZIO_WRITE_PIPELINE; | |
3184 | zio->io_bp_override = NULL; | |
3185 | BP_ZERO(bp); | |
3186 | ddt_exit(ddt); | |
62840030 | 3187 | return (zio); |
428870ff BB |
3188 | } |
3189 | ||
a6255b7f | 3190 | dio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3191 | zio->io_orig_size, zio->io_orig_size, &czp, NULL, NULL, |
bc77ba73 | 3192 | NULL, zio_ddt_ditto_write_done, dde, zio->io_priority, |
428870ff BB |
3193 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); |
3194 | ||
a6255b7f | 3195 | zio_push_transform(dio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3196 | dde->dde_lead_zio[DDT_PHYS_DITTO] = dio; |
3197 | } | |
3198 | ||
3199 | if (ddp->ddp_phys_birth != 0 || dde->dde_lead_zio[p] != NULL) { | |
3200 | if (ddp->ddp_phys_birth != 0) | |
3201 | ddt_bp_fill(ddp, bp, txg); | |
3202 | if (dde->dde_lead_zio[p] != NULL) | |
3203 | zio_add_child(zio, dde->dde_lead_zio[p]); | |
3204 | else | |
3205 | ddt_phys_addref(ddp); | |
3206 | } else if (zio->io_bp_override) { | |
3207 | ASSERT(bp->blk_birth == txg); | |
3208 | ASSERT(BP_EQUAL(bp, zio->io_bp_override)); | |
3209 | ddt_phys_fill(ddp, bp); | |
3210 | ddt_phys_addref(ddp); | |
3211 | } else { | |
a6255b7f | 3212 | cio = zio_write(zio, spa, txg, bp, zio->io_orig_abd, |
2aa34383 | 3213 | zio->io_orig_size, zio->io_orig_size, zp, |
bc77ba73 | 3214 | zio_ddt_child_write_ready, NULL, NULL, |
428870ff BB |
3215 | zio_ddt_child_write_done, dde, zio->io_priority, |
3216 | ZIO_DDT_CHILD_FLAGS(zio), &zio->io_bookmark); | |
3217 | ||
a6255b7f | 3218 | zio_push_transform(cio, zio->io_abd, zio->io_size, 0, NULL); |
428870ff BB |
3219 | dde->dde_lead_zio[p] = cio; |
3220 | } | |
3221 | ||
3222 | ddt_exit(ddt); | |
3223 | ||
3224 | if (cio) | |
3225 | zio_nowait(cio); | |
3226 | if (dio) | |
3227 | zio_nowait(dio); | |
3228 | ||
62840030 | 3229 | return (zio); |
428870ff BB |
3230 | } |
3231 | ||
3232 | ddt_entry_t *freedde; /* for debugging */ | |
b128c09f | 3233 | |
62840030 | 3234 | static zio_t * |
428870ff BB |
3235 | zio_ddt_free(zio_t *zio) |
3236 | { | |
3237 | spa_t *spa = zio->io_spa; | |
3238 | blkptr_t *bp = zio->io_bp; | |
3239 | ddt_t *ddt = ddt_select(spa, bp); | |
3240 | ddt_entry_t *dde; | |
3241 | ddt_phys_t *ddp; | |
3242 | ||
3243 | ASSERT(BP_GET_DEDUP(bp)); | |
3244 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
3245 | ||
3246 | ddt_enter(ddt); | |
3247 | freedde = dde = ddt_lookup(ddt, bp, B_TRUE); | |
5dc6af0e BB |
3248 | if (dde) { |
3249 | ddp = ddt_phys_select(dde, bp); | |
3250 | if (ddp) | |
3251 | ddt_phys_decref(ddp); | |
3252 | } | |
428870ff BB |
3253 | ddt_exit(ddt); |
3254 | ||
62840030 | 3255 | return (zio); |
428870ff BB |
3256 | } |
3257 | ||
3258 | /* | |
3259 | * ========================================================================== | |
3260 | * Allocate and free blocks | |
3261 | * ========================================================================== | |
3262 | */ | |
3dfb57a3 DB |
3263 | |
3264 | static zio_t * | |
492f64e9 | 3265 | zio_io_to_allocate(spa_t *spa, int allocator) |
3dfb57a3 DB |
3266 | { |
3267 | zio_t *zio; | |
3268 | ||
492f64e9 | 3269 | ASSERT(MUTEX_HELD(&spa->spa_alloc_locks[allocator])); |
3dfb57a3 | 3270 | |
492f64e9 | 3271 | zio = avl_first(&spa->spa_alloc_trees[allocator]); |
3dfb57a3 DB |
3272 | if (zio == NULL) |
3273 | return (NULL); | |
3274 | ||
3275 | ASSERT(IO_IS_ALLOCATING(zio)); | |
3276 | ||
3277 | /* | |
3278 | * Try to place a reservation for this zio. If we're unable to | |
3279 | * reserve then we throttle. | |
3280 | */ | |
492f64e9 | 3281 | ASSERT3U(zio->io_allocator, ==, allocator); |
cc99f275 | 3282 | if (!metaslab_class_throttle_reserve(zio->io_metaslab_class, |
492f64e9 | 3283 | zio->io_prop.zp_copies, zio->io_allocator, zio, 0)) { |
3dfb57a3 DB |
3284 | return (NULL); |
3285 | } | |
3286 | ||
492f64e9 | 3287 | avl_remove(&spa->spa_alloc_trees[allocator], zio); |
3dfb57a3 DB |
3288 | ASSERT3U(zio->io_stage, <, ZIO_STAGE_DVA_ALLOCATE); |
3289 | ||
3290 | return (zio); | |
3291 | } | |
3292 | ||
62840030 | 3293 | static zio_t * |
3dfb57a3 DB |
3294 | zio_dva_throttle(zio_t *zio) |
3295 | { | |
3296 | spa_t *spa = zio->io_spa; | |
3297 | zio_t *nio; | |
cc99f275 DB |
3298 | metaslab_class_t *mc; |
3299 | ||
3300 | /* locate an appropriate allocation class */ | |
3301 | mc = spa_preferred_class(spa, zio->io_size, zio->io_prop.zp_type, | |
3302 | zio->io_prop.zp_level, zio->io_prop.zp_zpl_smallblk); | |
3dfb57a3 DB |
3303 | |
3304 | if (zio->io_priority == ZIO_PRIORITY_SYNC_WRITE || | |
cc99f275 | 3305 | !mc->mc_alloc_throttle_enabled || |
3dfb57a3 DB |
3306 | zio->io_child_type == ZIO_CHILD_GANG || |
3307 | zio->io_flags & ZIO_FLAG_NODATA) { | |
62840030 | 3308 | return (zio); |
3dfb57a3 DB |
3309 | } |
3310 | ||
3311 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3312 | ||
3313 | ASSERT3U(zio->io_queued_timestamp, >, 0); | |
3314 | ASSERT(zio->io_stage == ZIO_STAGE_DVA_THROTTLE); | |
3315 | ||
492f64e9 PD |
3316 | zbookmark_phys_t *bm = &zio->io_bookmark; |
3317 | /* | |
3318 | * We want to try to use as many allocators as possible to help improve | |
3319 | * performance, but we also want logically adjacent IOs to be physically | |
3320 | * adjacent to improve sequential read performance. We chunk each object | |
3321 | * into 2^20 block regions, and then hash based on the objset, object, | |
3322 | * level, and region to accomplish both of these goals. | |
3323 | */ | |
3324 | zio->io_allocator = cityhash4(bm->zb_objset, bm->zb_object, | |
3325 | bm->zb_level, bm->zb_blkid >> 20) % spa->spa_alloc_count; | |
3326 | mutex_enter(&spa->spa_alloc_locks[zio->io_allocator]); | |
3dfb57a3 | 3327 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); |
cc99f275 | 3328 | zio->io_metaslab_class = mc; |
492f64e9 | 3329 | avl_add(&spa->spa_alloc_trees[zio->io_allocator], zio); |
cc99f275 | 3330 | nio = zio_io_to_allocate(spa, zio->io_allocator); |
492f64e9 | 3331 | mutex_exit(&spa->spa_alloc_locks[zio->io_allocator]); |
62840030 | 3332 | return (nio); |
3dfb57a3 DB |
3333 | } |
3334 | ||
cc99f275 | 3335 | static void |
492f64e9 | 3336 | zio_allocate_dispatch(spa_t *spa, int allocator) |
3dfb57a3 DB |
3337 | { |
3338 | zio_t *zio; | |
3339 | ||
492f64e9 PD |
3340 | mutex_enter(&spa->spa_alloc_locks[allocator]); |
3341 | zio = zio_io_to_allocate(spa, allocator); | |
3342 | mutex_exit(&spa->spa_alloc_locks[allocator]); | |
3dfb57a3 DB |
3343 | if (zio == NULL) |
3344 | return; | |
3345 | ||
3346 | ASSERT3U(zio->io_stage, ==, ZIO_STAGE_DVA_THROTTLE); | |
3347 | ASSERT0(zio->io_error); | |
3348 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, B_TRUE); | |
3349 | } | |
3350 | ||
62840030 | 3351 | static zio_t * |
34dc7c2f BB |
3352 | zio_dva_allocate(zio_t *zio) |
3353 | { | |
3354 | spa_t *spa = zio->io_spa; | |
cc99f275 | 3355 | metaslab_class_t *mc; |
34dc7c2f BB |
3356 | blkptr_t *bp = zio->io_bp; |
3357 | int error; | |
6d974228 | 3358 | int flags = 0; |
34dc7c2f | 3359 | |
9babb374 BB |
3360 | if (zio->io_gang_leader == NULL) { |
3361 | ASSERT(zio->io_child_type > ZIO_CHILD_GANG); | |
3362 | zio->io_gang_leader = zio; | |
3363 | } | |
3364 | ||
34dc7c2f | 3365 | ASSERT(BP_IS_HOLE(bp)); |
c99c9001 | 3366 | ASSERT0(BP_GET_NDVAS(bp)); |
428870ff BB |
3367 | ASSERT3U(zio->io_prop.zp_copies, >, 0); |
3368 | ASSERT3U(zio->io_prop.zp_copies, <=, spa_max_replication(spa)); | |
34dc7c2f BB |
3369 | ASSERT3U(zio->io_size, ==, BP_GET_PSIZE(bp)); |
3370 | ||
920dd524 | 3371 | flags |= (zio->io_flags & ZIO_FLAG_FASTWRITE) ? METASLAB_FASTWRITE : 0; |
3dfb57a3 DB |
3372 | if (zio->io_flags & ZIO_FLAG_NODATA) |
3373 | flags |= METASLAB_DONT_THROTTLE; | |
3374 | if (zio->io_flags & ZIO_FLAG_GANG_CHILD) | |
3375 | flags |= METASLAB_GANG_CHILD; | |
3376 | if (zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE) | |
3377 | flags |= METASLAB_ASYNC_ALLOC; | |
3378 | ||
cc99f275 DB |
3379 | /* |
3380 | * if not already chosen, locate an appropriate allocation class | |
3381 | */ | |
3382 | mc = zio->io_metaslab_class; | |
3383 | if (mc == NULL) { | |
3384 | mc = spa_preferred_class(spa, zio->io_size, | |
3385 | zio->io_prop.zp_type, zio->io_prop.zp_level, | |
3386 | zio->io_prop.zp_zpl_smallblk); | |
3387 | zio->io_metaslab_class = mc; | |
3388 | } | |
3389 | ||
b128c09f | 3390 | error = metaslab_alloc(spa, mc, zio->io_size, bp, |
4e21fd06 | 3391 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, |
492f64e9 | 3392 | &zio->io_alloc_list, zio, zio->io_allocator); |
34dc7c2f | 3393 | |
cc99f275 DB |
3394 | /* |
3395 | * Fallback to normal class when an alloc class is full | |
3396 | */ | |
3397 | if (error == ENOSPC && mc != spa_normal_class(spa)) { | |
3398 | /* | |
3399 | * If throttling, transfer reservation over to normal class. | |
3400 | * The io_allocator slot can remain the same even though we | |
3401 | * are switching classes. | |
3402 | */ | |
3403 | if (mc->mc_alloc_throttle_enabled && | |
3404 | (zio->io_flags & ZIO_FLAG_IO_ALLOCATING)) { | |
3405 | metaslab_class_throttle_unreserve(mc, | |
3406 | zio->io_prop.zp_copies, zio->io_allocator, zio); | |
3407 | zio->io_flags &= ~ZIO_FLAG_IO_ALLOCATING; | |
3408 | ||
3409 | mc = spa_normal_class(spa); | |
3410 | VERIFY(metaslab_class_throttle_reserve(mc, | |
3411 | zio->io_prop.zp_copies, zio->io_allocator, zio, | |
3412 | flags | METASLAB_MUST_RESERVE)); | |
3413 | } else { | |
3414 | mc = spa_normal_class(spa); | |
3415 | } | |
3416 | zio->io_metaslab_class = mc; | |
3417 | ||
3418 | error = metaslab_alloc(spa, mc, zio->io_size, bp, | |
3419 | zio->io_prop.zp_copies, zio->io_txg, NULL, flags, | |
3420 | &zio->io_alloc_list, zio, zio->io_allocator); | |
3421 | } | |
3422 | ||
3dfb57a3 | 3423 | if (error != 0) { |
964c2d69 | 3424 | zfs_dbgmsg("%s: metaslab allocation failure: zio %p, " |
6d974228 GW |
3425 | "size %llu, error %d", spa_name(spa), zio, zio->io_size, |
3426 | error); | |
b128c09f BB |
3427 | if (error == ENOSPC && zio->io_size > SPA_MINBLOCKSIZE) |
3428 | return (zio_write_gang_block(zio)); | |
34dc7c2f BB |
3429 | zio->io_error = error; |
3430 | } | |
3431 | ||
62840030 | 3432 | return (zio); |
34dc7c2f BB |
3433 | } |
3434 | ||
62840030 | 3435 | static zio_t * |
34dc7c2f BB |
3436 | zio_dva_free(zio_t *zio) |
3437 | { | |
b128c09f | 3438 | metaslab_free(zio->io_spa, zio->io_bp, zio->io_txg, B_FALSE); |
34dc7c2f | 3439 | |
62840030 | 3440 | return (zio); |
34dc7c2f BB |
3441 | } |
3442 | ||
62840030 | 3443 | static zio_t * |
34dc7c2f BB |
3444 | zio_dva_claim(zio_t *zio) |
3445 | { | |
b128c09f BB |
3446 | int error; |
3447 | ||
3448 | error = metaslab_claim(zio->io_spa, zio->io_bp, zio->io_txg); | |
3449 | if (error) | |
3450 | zio->io_error = error; | |
34dc7c2f | 3451 | |
62840030 | 3452 | return (zio); |
34dc7c2f BB |
3453 | } |
3454 | ||
b128c09f BB |
3455 | /* |
3456 | * Undo an allocation. This is used by zio_done() when an I/O fails | |
3457 | * and we want to give back the block we just allocated. | |
3458 | * This handles both normal blocks and gang blocks. | |
3459 | */ | |
3460 | static void | |
3461 | zio_dva_unallocate(zio_t *zio, zio_gang_node_t *gn, blkptr_t *bp) | |
3462 | { | |
b128c09f | 3463 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp)); |
428870ff | 3464 | ASSERT(zio->io_bp_override == NULL); |
b128c09f BB |
3465 | |
3466 | if (!BP_IS_HOLE(bp)) | |
428870ff | 3467 | metaslab_free(zio->io_spa, bp, bp->blk_birth, B_TRUE); |
b128c09f BB |
3468 | |
3469 | if (gn != NULL) { | |
1c27024e | 3470 | for (int g = 0; g < SPA_GBH_NBLKPTRS; g++) { |
b128c09f BB |
3471 | zio_dva_unallocate(zio, gn->gn_child[g], |
3472 | &gn->gn_gbh->zg_blkptr[g]); | |
3473 | } | |
3474 | } | |
3475 | } | |
3476 | ||
3477 | /* | |
3478 | * Try to allocate an intent log block. Return 0 on success, errno on failure. | |
3479 | */ | |
3480 | int | |
b5256303 TC |
3481 | zio_alloc_zil(spa_t *spa, objset_t *os, uint64_t txg, blkptr_t *new_bp, |
3482 | uint64_t size, boolean_t *slog) | |
b128c09f | 3483 | { |
428870ff | 3484 | int error = 1; |
4e21fd06 | 3485 | zio_alloc_list_t io_alloc_list; |
b128c09f | 3486 | |
428870ff BB |
3487 | ASSERT(txg > spa_syncing_txg(spa)); |
3488 | ||
4e21fd06 | 3489 | metaslab_trace_init(&io_alloc_list); |
cc99f275 DB |
3490 | |
3491 | /* | |
3492 | * Block pointer fields are useful to metaslabs for stats and debugging. | |
3493 | * Fill in the obvious ones before calling into metaslab_alloc(). | |
3494 | */ | |
3495 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); | |
3496 | BP_SET_PSIZE(new_bp, size); | |
3497 | BP_SET_LEVEL(new_bp, 0); | |
3498 | ||
492f64e9 PD |
3499 | /* |
3500 | * When allocating a zil block, we don't have information about | |
3501 | * the final destination of the block except the objset it's part | |
3502 | * of, so we just hash the objset ID to pick the allocator to get | |
3503 | * some parallelism. | |
3504 | */ | |
1b7c1e5c | 3505 | error = metaslab_alloc(spa, spa_log_class(spa), size, new_bp, 1, |
492f64e9 PD |
3506 | txg, NULL, METASLAB_FASTWRITE, &io_alloc_list, NULL, |
3507 | cityhash4(0, 0, 0, os->os_dsl_dataset->ds_object) % | |
3508 | spa->spa_alloc_count); | |
1b7c1e5c GDN |
3509 | if (error == 0) { |
3510 | *slog = TRUE; | |
3511 | } else { | |
428870ff | 3512 | error = metaslab_alloc(spa, spa_normal_class(spa), size, |
4e21fd06 | 3513 | new_bp, 1, txg, NULL, METASLAB_FASTWRITE, |
492f64e9 PD |
3514 | &io_alloc_list, NULL, cityhash4(0, 0, 0, |
3515 | os->os_dsl_dataset->ds_object) % spa->spa_alloc_count); | |
1b7c1e5c GDN |
3516 | if (error == 0) |
3517 | *slog = FALSE; | |
ebf8e3a2 | 3518 | } |
4e21fd06 | 3519 | metaslab_trace_fini(&io_alloc_list); |
b128c09f BB |
3520 | |
3521 | if (error == 0) { | |
3522 | BP_SET_LSIZE(new_bp, size); | |
3523 | BP_SET_PSIZE(new_bp, size); | |
3524 | BP_SET_COMPRESS(new_bp, ZIO_COMPRESS_OFF); | |
428870ff BB |
3525 | BP_SET_CHECKSUM(new_bp, |
3526 | spa_version(spa) >= SPA_VERSION_SLIM_ZIL | |
3527 | ? ZIO_CHECKSUM_ZILOG2 : ZIO_CHECKSUM_ZILOG); | |
b128c09f BB |
3528 | BP_SET_TYPE(new_bp, DMU_OT_INTENT_LOG); |
3529 | BP_SET_LEVEL(new_bp, 0); | |
428870ff | 3530 | BP_SET_DEDUP(new_bp, 0); |
b128c09f | 3531 | BP_SET_BYTEORDER(new_bp, ZFS_HOST_BYTEORDER); |
b5256303 TC |
3532 | |
3533 | /* | |
3534 | * encrypted blocks will require an IV and salt. We generate | |
3535 | * these now since we will not be rewriting the bp at | |
3536 | * rewrite time. | |
3537 | */ | |
3538 | if (os->os_encrypted) { | |
3539 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3540 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3541 | ||
3542 | BP_SET_CRYPT(new_bp, B_TRUE); | |
3543 | VERIFY0(spa_crypt_get_salt(spa, | |
3544 | dmu_objset_id(os), salt)); | |
3545 | VERIFY0(zio_crypt_generate_iv(iv)); | |
3546 | ||
3547 | zio_crypt_encode_params_bp(new_bp, salt, iv); | |
3548 | } | |
1ce23dca PS |
3549 | } else { |
3550 | zfs_dbgmsg("%s: zil block allocation failure: " | |
3551 | "size %llu, error %d", spa_name(spa), size, error); | |
b128c09f BB |
3552 | } |
3553 | ||
3554 | return (error); | |
3555 | } | |
3556 | ||
34dc7c2f BB |
3557 | /* |
3558 | * ========================================================================== | |
3559 | * Read and write to physical devices | |
3560 | * ========================================================================== | |
3561 | */ | |
98b25418 GW |
3562 | |
3563 | ||
3564 | /* | |
3565 | * Issue an I/O to the underlying vdev. Typically the issue pipeline | |
3566 | * stops after this stage and will resume upon I/O completion. | |
3567 | * However, there are instances where the vdev layer may need to | |
3568 | * continue the pipeline when an I/O was not issued. Since the I/O | |
3569 | * that was sent to the vdev layer might be different than the one | |
3570 | * currently active in the pipeline (see vdev_queue_io()), we explicitly | |
3571 | * force the underlying vdev layers to call either zio_execute() or | |
3572 | * zio_interrupt() to ensure that the pipeline continues with the correct I/O. | |
3573 | */ | |
62840030 | 3574 | static zio_t * |
34dc7c2f BB |
3575 | zio_vdev_io_start(zio_t *zio) |
3576 | { | |
3577 | vdev_t *vd = zio->io_vd; | |
34dc7c2f BB |
3578 | uint64_t align; |
3579 | spa_t *spa = zio->io_spa; | |
3580 | ||
193a37cb TH |
3581 | zio->io_delay = 0; |
3582 | ||
b128c09f BB |
3583 | ASSERT(zio->io_error == 0); |
3584 | ASSERT(zio->io_child_error[ZIO_CHILD_VDEV] == 0); | |
34dc7c2f | 3585 | |
b128c09f BB |
3586 | if (vd == NULL) { |
3587 | if (!(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3588 | spa_config_enter(spa, SCL_ZIO, zio, RW_READER); | |
34dc7c2f | 3589 | |
b128c09f BB |
3590 | /* |
3591 | * The mirror_ops handle multiple DVAs in a single BP. | |
3592 | */ | |
98b25418 | 3593 | vdev_mirror_ops.vdev_op_io_start(zio); |
62840030 | 3594 | return (NULL); |
34dc7c2f BB |
3595 | } |
3596 | ||
3dfb57a3 | 3597 | ASSERT3P(zio->io_logical, !=, zio); |
6cb8e530 PZ |
3598 | if (zio->io_type == ZIO_TYPE_WRITE) { |
3599 | ASSERT(spa->spa_trust_config); | |
3600 | ||
a1d477c2 MA |
3601 | /* |
3602 | * Note: the code can handle other kinds of writes, | |
3603 | * but we don't expect them. | |
3604 | */ | |
6cb8e530 PZ |
3605 | if (zio->io_vd->vdev_removing) { |
3606 | ASSERT(zio->io_flags & | |
3607 | (ZIO_FLAG_PHYSICAL | ZIO_FLAG_SELF_HEAL | | |
3608 | ZIO_FLAG_RESILVER | ZIO_FLAG_INDUCE_DAMAGE)); | |
3609 | } | |
a1d477c2 | 3610 | } |
3dfb57a3 | 3611 | |
b128c09f BB |
3612 | align = 1ULL << vd->vdev_top->vdev_ashift; |
3613 | ||
b02fe35d AR |
3614 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL) && |
3615 | P2PHASE(zio->io_size, align) != 0) { | |
3616 | /* Transform logical writes to be a full physical block size. */ | |
34dc7c2f | 3617 | uint64_t asize = P2ROUNDUP(zio->io_size, align); |
a6255b7f | 3618 | abd_t *abuf = abd_alloc_sametype(zio->io_abd, asize); |
178e73b3 | 3619 | ASSERT(vd == vd->vdev_top); |
34dc7c2f | 3620 | if (zio->io_type == ZIO_TYPE_WRITE) { |
a6255b7f DQ |
3621 | abd_copy(abuf, zio->io_abd, zio->io_size); |
3622 | abd_zero_off(abuf, zio->io_size, asize - zio->io_size); | |
34dc7c2f | 3623 | } |
b128c09f | 3624 | zio_push_transform(zio, abuf, asize, asize, zio_subblock); |
34dc7c2f BB |
3625 | } |
3626 | ||
b02fe35d AR |
3627 | /* |
3628 | * If this is not a physical io, make sure that it is properly aligned | |
3629 | * before proceeding. | |
3630 | */ | |
3631 | if (!(zio->io_flags & ZIO_FLAG_PHYSICAL)) { | |
3632 | ASSERT0(P2PHASE(zio->io_offset, align)); | |
3633 | ASSERT0(P2PHASE(zio->io_size, align)); | |
3634 | } else { | |
3635 | /* | |
3636 | * For physical writes, we allow 512b aligned writes and assume | |
3637 | * the device will perform a read-modify-write as necessary. | |
3638 | */ | |
3639 | ASSERT0(P2PHASE(zio->io_offset, SPA_MINBLOCKSIZE)); | |
3640 | ASSERT0(P2PHASE(zio->io_size, SPA_MINBLOCKSIZE)); | |
3641 | } | |
3642 | ||
572e2857 | 3643 | VERIFY(zio->io_type != ZIO_TYPE_WRITE || spa_writeable(spa)); |
fb5f0bc8 BB |
3644 | |
3645 | /* | |
3646 | * If this is a repair I/O, and there's no self-healing involved -- | |
3647 | * that is, we're just resilvering what we expect to resilver -- | |
3648 | * then don't do the I/O unless zio's txg is actually in vd's DTL. | |
9e052db4 MA |
3649 | * This prevents spurious resilvering. |
3650 | * | |
3651 | * There are a few ways that we can end up creating these spurious | |
3652 | * resilver i/os: | |
3653 | * | |
3654 | * 1. A resilver i/o will be issued if any DVA in the BP has a | |
3655 | * dirty DTL. The mirror code will issue resilver writes to | |
3656 | * each DVA, including the one(s) that are not on vdevs with dirty | |
3657 | * DTLs. | |
3658 | * | |
3659 | * 2. With nested replication, which happens when we have a | |
3660 | * "replacing" or "spare" vdev that's a child of a mirror or raidz. | |
3661 | * For example, given mirror(replacing(A+B), C), it's likely that | |
3662 | * only A is out of date (it's the new device). In this case, we'll | |
3663 | * read from C, then use the data to resilver A+B -- but we don't | |
3664 | * actually want to resilver B, just A. The top-level mirror has no | |
3665 | * way to know this, so instead we just discard unnecessary repairs | |
3666 | * as we work our way down the vdev tree. | |
3667 | * | |
3668 | * 3. ZTEST also creates mirrors of mirrors, mirrors of raidz, etc. | |
3669 | * The same logic applies to any form of nested replication: ditto | |
3670 | * + mirror, RAID-Z + replacing, etc. | |
3671 | * | |
3672 | * However, indirect vdevs point off to other vdevs which may have | |
3673 | * DTL's, so we never bypass them. The child i/os on concrete vdevs | |
3674 | * will be properly bypassed instead. | |
fb5f0bc8 BB |
3675 | */ |
3676 | if ((zio->io_flags & ZIO_FLAG_IO_REPAIR) && | |
3677 | !(zio->io_flags & ZIO_FLAG_SELF_HEAL) && | |
3678 | zio->io_txg != 0 && /* not a delegated i/o */ | |
9e052db4 | 3679 | vd->vdev_ops != &vdev_indirect_ops && |
fb5f0bc8 BB |
3680 | !vdev_dtl_contains(vd, DTL_PARTIAL, zio->io_txg, 1)) { |
3681 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
fb5f0bc8 | 3682 | zio_vdev_io_bypass(zio); |
62840030 | 3683 | return (zio); |
fb5f0bc8 | 3684 | } |
34dc7c2f | 3685 | |
b128c09f BB |
3686 | if (vd->vdev_ops->vdev_op_leaf && |
3687 | (zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE)) { | |
3688 | ||
b0bc7a84 | 3689 | if (zio->io_type == ZIO_TYPE_READ && vdev_cache_read(zio)) |
62840030 | 3690 | return (zio); |
b128c09f BB |
3691 | |
3692 | if ((zio = vdev_queue_io(zio)) == NULL) | |
62840030 | 3693 | return (NULL); |
b128c09f BB |
3694 | |
3695 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3696 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f | 3697 | zio_interrupt(zio); |
62840030 | 3698 | return (NULL); |
b128c09f | 3699 | } |
67103816 | 3700 | zio->io_delay = gethrtime(); |
b128c09f BB |
3701 | } |
3702 | ||
98b25418 | 3703 | vd->vdev_ops->vdev_op_io_start(zio); |
62840030 | 3704 | return (NULL); |
34dc7c2f BB |
3705 | } |
3706 | ||
62840030 | 3707 | static zio_t * |
34dc7c2f BB |
3708 | zio_vdev_io_done(zio_t *zio) |
3709 | { | |
b128c09f BB |
3710 | vdev_t *vd = zio->io_vd; |
3711 | vdev_ops_t *ops = vd ? vd->vdev_ops : &vdev_mirror_ops; | |
3712 | boolean_t unexpected_error = B_FALSE; | |
34dc7c2f | 3713 | |
ddc751d5 | 3714 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3715 | return (NULL); |
ddc751d5 | 3716 | } |
34dc7c2f | 3717 | |
b128c09f BB |
3718 | ASSERT(zio->io_type == ZIO_TYPE_READ || zio->io_type == ZIO_TYPE_WRITE); |
3719 | ||
193a37cb TH |
3720 | if (zio->io_delay) |
3721 | zio->io_delay = gethrtime() - zio->io_delay; | |
3722 | ||
b128c09f BB |
3723 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf) { |
3724 | ||
3725 | vdev_queue_io_done(zio); | |
3726 | ||
3727 | if (zio->io_type == ZIO_TYPE_WRITE) | |
3728 | vdev_cache_write(zio); | |
3729 | ||
3730 | if (zio_injection_enabled && zio->io_error == 0) | |
d977122d DB |
3731 | zio->io_error = zio_handle_device_injections(vd, zio, |
3732 | EIO, EILSEQ); | |
b128c09f BB |
3733 | |
3734 | if (zio_injection_enabled && zio->io_error == 0) | |
3735 | zio->io_error = zio_handle_label_injection(zio, EIO); | |
3736 | ||
3737 | if (zio->io_error) { | |
3738 | if (!vdev_accessible(vd, zio)) { | |
2e528b49 | 3739 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3740 | } else { |
3741 | unexpected_error = B_TRUE; | |
3742 | } | |
3743 | } | |
3744 | } | |
3745 | ||
3746 | ops->vdev_op_io_done(zio); | |
34dc7c2f | 3747 | |
f43615d0 | 3748 | if (unexpected_error) |
d164b209 | 3749 | VERIFY(vdev_probe(vd, zio) == NULL); |
34dc7c2f | 3750 | |
62840030 | 3751 | return (zio); |
34dc7c2f BB |
3752 | } |
3753 | ||
a8b2e306 TC |
3754 | /* |
3755 | * This function is used to change the priority of an existing zio that is | |
3756 | * currently in-flight. This is used by the arc to upgrade priority in the | |
3757 | * event that a demand read is made for a block that is currently queued | |
3758 | * as a scrub or async read IO. Otherwise, the high priority read request | |
3759 | * would end up having to wait for the lower priority IO. | |
3760 | */ | |
3761 | void | |
3762 | zio_change_priority(zio_t *pio, zio_priority_t priority) | |
3763 | { | |
3764 | zio_t *cio, *cio_next; | |
3765 | zio_link_t *zl = NULL; | |
3766 | ||
3767 | ASSERT3U(priority, <, ZIO_PRIORITY_NUM_QUEUEABLE); | |
3768 | ||
3769 | if (pio->io_vd != NULL && pio->io_vd->vdev_ops->vdev_op_leaf) { | |
3770 | vdev_queue_change_io_priority(pio, priority); | |
3771 | } else { | |
3772 | pio->io_priority = priority; | |
3773 | } | |
3774 | ||
3775 | mutex_enter(&pio->io_lock); | |
3776 | for (cio = zio_walk_children(pio, &zl); cio != NULL; cio = cio_next) { | |
3777 | cio_next = zio_walk_children(pio, &zl); | |
3778 | zio_change_priority(cio, priority); | |
3779 | } | |
3780 | mutex_exit(&pio->io_lock); | |
3781 | } | |
3782 | ||
428870ff BB |
3783 | /* |
3784 | * For non-raidz ZIOs, we can just copy aside the bad data read from the | |
3785 | * disk, and use that to finish the checksum ereport later. | |
3786 | */ | |
3787 | static void | |
3788 | zio_vsd_default_cksum_finish(zio_cksum_report_t *zcr, | |
84c07ada | 3789 | const abd_t *good_buf) |
428870ff BB |
3790 | { |
3791 | /* no processing needed */ | |
3792 | zfs_ereport_finish_checksum(zcr, good_buf, zcr->zcr_cbdata, B_FALSE); | |
3793 | } | |
3794 | ||
3795 | /*ARGSUSED*/ | |
3796 | void | |
3797 | zio_vsd_default_cksum_report(zio_t *zio, zio_cksum_report_t *zcr, void *ignored) | |
3798 | { | |
84c07ada | 3799 | void *abd = abd_alloc_sametype(zio->io_abd, zio->io_size); |
428870ff | 3800 | |
84c07ada | 3801 | abd_copy(abd, zio->io_abd, zio->io_size); |
428870ff BB |
3802 | |
3803 | zcr->zcr_cbinfo = zio->io_size; | |
84c07ada | 3804 | zcr->zcr_cbdata = abd; |
428870ff | 3805 | zcr->zcr_finish = zio_vsd_default_cksum_finish; |
84c07ada | 3806 | zcr->zcr_free = zio_abd_free; |
428870ff BB |
3807 | } |
3808 | ||
62840030 | 3809 | static zio_t * |
34dc7c2f BB |
3810 | zio_vdev_io_assess(zio_t *zio) |
3811 | { | |
3812 | vdev_t *vd = zio->io_vd; | |
b128c09f | 3813 | |
ddc751d5 | 3814 | if (zio_wait_for_children(zio, ZIO_CHILD_VDEV_BIT, ZIO_WAIT_DONE)) { |
62840030 | 3815 | return (NULL); |
ddc751d5 | 3816 | } |
b128c09f BB |
3817 | |
3818 | if (vd == NULL && !(zio->io_flags & ZIO_FLAG_CONFIG_WRITER)) | |
3819 | spa_config_exit(zio->io_spa, SCL_ZIO, zio); | |
3820 | ||
3821 | if (zio->io_vsd != NULL) { | |
428870ff | 3822 | zio->io_vsd_ops->vsd_free(zio); |
b128c09f | 3823 | zio->io_vsd = NULL; |
34dc7c2f BB |
3824 | } |
3825 | ||
b128c09f | 3826 | if (zio_injection_enabled && zio->io_error == 0) |
34dc7c2f BB |
3827 | zio->io_error = zio_handle_fault_injection(zio, EIO); |
3828 | ||
3829 | /* | |
3830 | * If the I/O failed, determine whether we should attempt to retry it. | |
428870ff BB |
3831 | * |
3832 | * On retry, we cut in line in the issue queue, since we don't want | |
3833 | * compression/checksumming/etc. work to prevent our (cheap) IO reissue. | |
34dc7c2f | 3834 | */ |
b128c09f BB |
3835 | if (zio->io_error && vd == NULL && |
3836 | !(zio->io_flags & (ZIO_FLAG_DONT_RETRY | ZIO_FLAG_IO_RETRY))) { | |
3837 | ASSERT(!(zio->io_flags & ZIO_FLAG_DONT_QUEUE)); /* not a leaf */ | |
3838 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_BYPASS)); /* not a leaf */ | |
34dc7c2f | 3839 | zio->io_error = 0; |
b128c09f BB |
3840 | zio->io_flags |= ZIO_FLAG_IO_RETRY | |
3841 | ZIO_FLAG_DONT_CACHE | ZIO_FLAG_DONT_AGGREGATE; | |
428870ff BB |
3842 | zio->io_stage = ZIO_STAGE_VDEV_IO_START >> 1; |
3843 | zio_taskq_dispatch(zio, ZIO_TASKQ_ISSUE, | |
3844 | zio_requeue_io_start_cut_in_line); | |
62840030 | 3845 | return (NULL); |
34dc7c2f BB |
3846 | } |
3847 | ||
b128c09f BB |
3848 | /* |
3849 | * If we got an error on a leaf device, convert it to ENXIO | |
3850 | * if the device is not accessible at all. | |
3851 | */ | |
3852 | if (zio->io_error && vd != NULL && vd->vdev_ops->vdev_op_leaf && | |
3853 | !vdev_accessible(vd, zio)) | |
2e528b49 | 3854 | zio->io_error = SET_ERROR(ENXIO); |
b128c09f BB |
3855 | |
3856 | /* | |
3857 | * If we can't write to an interior vdev (mirror or RAID-Z), | |
3858 | * set vdev_cant_write so that we stop trying to allocate from it. | |
3859 | */ | |
3860 | if (zio->io_error == ENXIO && zio->io_type == ZIO_TYPE_WRITE && | |
13fe0198 | 3861 | vd != NULL && !vd->vdev_ops->vdev_op_leaf) { |
b128c09f | 3862 | vd->vdev_cant_write = B_TRUE; |
13fe0198 | 3863 | } |
b128c09f | 3864 | |
298ec40b GM |
3865 | /* |
3866 | * If a cache flush returns ENOTSUP or ENOTTY, we know that no future | |
3867 | * attempts will ever succeed. In this case we set a persistent bit so | |
3868 | * that we don't bother with it in the future. | |
3869 | */ | |
3870 | if ((zio->io_error == ENOTSUP || zio->io_error == ENOTTY) && | |
3871 | zio->io_type == ZIO_TYPE_IOCTL && | |
3872 | zio->io_cmd == DKIOCFLUSHWRITECACHE && vd != NULL) | |
3873 | vd->vdev_nowritecache = B_TRUE; | |
3874 | ||
b128c09f BB |
3875 | if (zio->io_error) |
3876 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; | |
3877 | ||
e8b96c60 MA |
3878 | if (vd != NULL && vd->vdev_ops->vdev_op_leaf && |
3879 | zio->io_physdone != NULL) { | |
3880 | ASSERT(!(zio->io_flags & ZIO_FLAG_DELEGATED)); | |
3881 | ASSERT(zio->io_child_type == ZIO_CHILD_VDEV); | |
3882 | zio->io_physdone(zio->io_logical); | |
3883 | } | |
3884 | ||
62840030 | 3885 | return (zio); |
34dc7c2f BB |
3886 | } |
3887 | ||
3888 | void | |
3889 | zio_vdev_io_reissue(zio_t *zio) | |
3890 | { | |
3891 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3892 | ASSERT(zio->io_error == 0); | |
3893 | ||
428870ff | 3894 | zio->io_stage >>= 1; |
34dc7c2f BB |
3895 | } |
3896 | ||
3897 | void | |
3898 | zio_vdev_io_redone(zio_t *zio) | |
3899 | { | |
3900 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_DONE); | |
3901 | ||
428870ff | 3902 | zio->io_stage >>= 1; |
34dc7c2f BB |
3903 | } |
3904 | ||
3905 | void | |
3906 | zio_vdev_io_bypass(zio_t *zio) | |
3907 | { | |
3908 | ASSERT(zio->io_stage == ZIO_STAGE_VDEV_IO_START); | |
3909 | ASSERT(zio->io_error == 0); | |
3910 | ||
3911 | zio->io_flags |= ZIO_FLAG_IO_BYPASS; | |
428870ff | 3912 | zio->io_stage = ZIO_STAGE_VDEV_IO_ASSESS >> 1; |
34dc7c2f BB |
3913 | } |
3914 | ||
b5256303 TC |
3915 | /* |
3916 | * ========================================================================== | |
3917 | * Encrypt and store encryption parameters | |
3918 | * ========================================================================== | |
3919 | */ | |
3920 | ||
3921 | ||
3922 | /* | |
3923 | * This function is used for ZIO_STAGE_ENCRYPT. It is responsible for | |
3924 | * managing the storage of encryption parameters and passing them to the | |
3925 | * lower-level encryption functions. | |
3926 | */ | |
62840030 | 3927 | static zio_t * |
b5256303 TC |
3928 | zio_encrypt(zio_t *zio) |
3929 | { | |
3930 | zio_prop_t *zp = &zio->io_prop; | |
3931 | spa_t *spa = zio->io_spa; | |
3932 | blkptr_t *bp = zio->io_bp; | |
3933 | uint64_t psize = BP_GET_PSIZE(bp); | |
ae76f45c | 3934 | uint64_t dsobj = zio->io_bookmark.zb_objset; |
b5256303 TC |
3935 | dmu_object_type_t ot = BP_GET_TYPE(bp); |
3936 | void *enc_buf = NULL; | |
3937 | abd_t *eabd = NULL; | |
3938 | uint8_t salt[ZIO_DATA_SALT_LEN]; | |
3939 | uint8_t iv[ZIO_DATA_IV_LEN]; | |
3940 | uint8_t mac[ZIO_DATA_MAC_LEN]; | |
3941 | boolean_t no_crypt = B_FALSE; | |
3942 | ||
3943 | /* the root zio already encrypted the data */ | |
3944 | if (zio->io_child_type == ZIO_CHILD_GANG) | |
62840030 | 3945 | return (zio); |
b5256303 TC |
3946 | |
3947 | /* only ZIL blocks are re-encrypted on rewrite */ | |
3948 | if (!IO_IS_ALLOCATING(zio) && ot != DMU_OT_INTENT_LOG) | |
62840030 | 3949 | return (zio); |
b5256303 TC |
3950 | |
3951 | if (!(zp->zp_encrypt || BP_IS_ENCRYPTED(bp))) { | |
3952 | BP_SET_CRYPT(bp, B_FALSE); | |
62840030 | 3953 | return (zio); |
b5256303 TC |
3954 | } |
3955 | ||
3956 | /* if we are doing raw encryption set the provided encryption params */ | |
3957 | if (zio->io_flags & ZIO_FLAG_RAW_ENCRYPT) { | |
ae76f45c | 3958 | ASSERT0(BP_GET_LEVEL(bp)); |
b5256303 TC |
3959 | BP_SET_CRYPT(bp, B_TRUE); |
3960 | BP_SET_BYTEORDER(bp, zp->zp_byteorder); | |
3961 | if (ot != DMU_OT_OBJSET) | |
3962 | zio_crypt_encode_mac_bp(bp, zp->zp_mac); | |
ae76f45c TC |
3963 | |
3964 | /* dnode blocks must be written out in the provided byteorder */ | |
3965 | if (zp->zp_byteorder != ZFS_HOST_BYTEORDER && | |
3966 | ot == DMU_OT_DNODE) { | |
3967 | void *bswap_buf = zio_buf_alloc(psize); | |
3968 | abd_t *babd = abd_get_from_buf(bswap_buf, psize); | |
3969 | ||
3970 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
3971 | abd_copy_to_buf(bswap_buf, zio->io_abd, psize); | |
3972 | dmu_ot_byteswap[DMU_OT_BYTESWAP(ot)].ob_func(bswap_buf, | |
3973 | psize); | |
3974 | ||
3975 | abd_take_ownership_of_buf(babd, B_TRUE); | |
3976 | zio_push_transform(zio, babd, psize, psize, NULL); | |
3977 | } | |
3978 | ||
b5256303 TC |
3979 | if (DMU_OT_IS_ENCRYPTED(ot)) |
3980 | zio_crypt_encode_params_bp(bp, zp->zp_salt, zp->zp_iv); | |
62840030 | 3981 | return (zio); |
b5256303 TC |
3982 | } |
3983 | ||
3984 | /* indirect blocks only maintain a cksum of the lower level MACs */ | |
3985 | if (BP_GET_LEVEL(bp) > 0) { | |
3986 | BP_SET_CRYPT(bp, B_TRUE); | |
3987 | VERIFY0(zio_crypt_do_indirect_mac_checksum_abd(B_TRUE, | |
3988 | zio->io_orig_abd, BP_GET_LSIZE(bp), BP_SHOULD_BYTESWAP(bp), | |
3989 | mac)); | |
3990 | zio_crypt_encode_mac_bp(bp, mac); | |
62840030 | 3991 | return (zio); |
b5256303 TC |
3992 | } |
3993 | ||
3994 | /* | |
3995 | * Objset blocks are a special case since they have 2 256-bit MACs | |
3996 | * embedded within them. | |
3997 | */ | |
3998 | if (ot == DMU_OT_OBJSET) { | |
3999 | ASSERT0(DMU_OT_IS_ENCRYPTED(ot)); | |
4000 | ASSERT3U(BP_GET_COMPRESS(bp), ==, ZIO_COMPRESS_OFF); | |
4001 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4002 | VERIFY0(spa_do_crypt_objset_mac_abd(B_TRUE, spa, dsobj, |
4003 | zio->io_abd, psize, BP_SHOULD_BYTESWAP(bp))); | |
62840030 | 4004 | return (zio); |
b5256303 TC |
4005 | } |
4006 | ||
4007 | /* unencrypted object types are only authenticated with a MAC */ | |
4008 | if (!DMU_OT_IS_ENCRYPTED(ot)) { | |
4009 | BP_SET_CRYPT(bp, B_TRUE); | |
ae76f45c TC |
4010 | VERIFY0(spa_do_crypt_mac_abd(B_TRUE, spa, dsobj, |
4011 | zio->io_abd, psize, mac)); | |
b5256303 | 4012 | zio_crypt_encode_mac_bp(bp, mac); |
62840030 | 4013 | return (zio); |
b5256303 TC |
4014 | } |
4015 | ||
4016 | /* | |
4017 | * Later passes of sync-to-convergence may decide to rewrite data | |
4018 | * in place to avoid more disk reallocations. This presents a problem | |
d611989f | 4019 | * for encryption because this constitutes rewriting the new data with |
b5256303 TC |
4020 | * the same encryption key and IV. However, this only applies to blocks |
4021 | * in the MOS (particularly the spacemaps) and we do not encrypt the | |
4022 | * MOS. We assert that the zio is allocating or an intent log write | |
4023 | * to enforce this. | |
4024 | */ | |
4025 | ASSERT(IO_IS_ALLOCATING(zio) || ot == DMU_OT_INTENT_LOG); | |
4026 | ASSERT(BP_GET_LEVEL(bp) == 0 || ot == DMU_OT_INTENT_LOG); | |
4027 | ASSERT(spa_feature_is_active(spa, SPA_FEATURE_ENCRYPTION)); | |
4028 | ASSERT3U(psize, !=, 0); | |
4029 | ||
4030 | enc_buf = zio_buf_alloc(psize); | |
4031 | eabd = abd_get_from_buf(enc_buf, psize); | |
4032 | abd_take_ownership_of_buf(eabd, B_TRUE); | |
4033 | ||
4034 | /* | |
4035 | * For an explanation of what encryption parameters are stored | |
4036 | * where, see the block comment in zio_crypt.c. | |
4037 | */ | |
4038 | if (ot == DMU_OT_INTENT_LOG) { | |
4039 | zio_crypt_decode_params_bp(bp, salt, iv); | |
4040 | } else { | |
4041 | BP_SET_CRYPT(bp, B_TRUE); | |
4042 | } | |
4043 | ||
4044 | /* Perform the encryption. This should not fail */ | |
be9a5c35 TC |
4045 | VERIFY0(spa_do_crypt_abd(B_TRUE, spa, &zio->io_bookmark, |
4046 | BP_GET_TYPE(bp), BP_GET_DEDUP(bp), BP_SHOULD_BYTESWAP(bp), | |
4047 | salt, iv, mac, psize, zio->io_abd, eabd, &no_crypt)); | |
b5256303 TC |
4048 | |
4049 | /* encode encryption metadata into the bp */ | |
4050 | if (ot == DMU_OT_INTENT_LOG) { | |
4051 | /* | |
4052 | * ZIL blocks store the MAC in the embedded checksum, so the | |
4053 | * transform must always be applied. | |
4054 | */ | |
4055 | zio_crypt_encode_mac_zil(enc_buf, mac); | |
4056 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4057 | } else { | |
4058 | BP_SET_CRYPT(bp, B_TRUE); | |
4059 | zio_crypt_encode_params_bp(bp, salt, iv); | |
4060 | zio_crypt_encode_mac_bp(bp, mac); | |
4061 | ||
4062 | if (no_crypt) { | |
4063 | ASSERT3U(ot, ==, DMU_OT_DNODE); | |
4064 | abd_free(eabd); | |
4065 | } else { | |
4066 | zio_push_transform(zio, eabd, psize, psize, NULL); | |
4067 | } | |
4068 | } | |
4069 | ||
62840030 | 4070 | return (zio); |
b5256303 TC |
4071 | } |
4072 | ||
34dc7c2f BB |
4073 | /* |
4074 | * ========================================================================== | |
4075 | * Generate and verify checksums | |
4076 | * ========================================================================== | |
4077 | */ | |
62840030 | 4078 | static zio_t * |
34dc7c2f BB |
4079 | zio_checksum_generate(zio_t *zio) |
4080 | { | |
34dc7c2f | 4081 | blkptr_t *bp = zio->io_bp; |
b128c09f | 4082 | enum zio_checksum checksum; |
34dc7c2f | 4083 | |
b128c09f BB |
4084 | if (bp == NULL) { |
4085 | /* | |
4086 | * This is zio_write_phys(). | |
4087 | * We're either generating a label checksum, or none at all. | |
4088 | */ | |
4089 | checksum = zio->io_prop.zp_checksum; | |
34dc7c2f | 4090 | |
b128c09f | 4091 | if (checksum == ZIO_CHECKSUM_OFF) |
62840030 | 4092 | return (zio); |
b128c09f BB |
4093 | |
4094 | ASSERT(checksum == ZIO_CHECKSUM_LABEL); | |
4095 | } else { | |
4096 | if (BP_IS_GANG(bp) && zio->io_child_type == ZIO_CHILD_GANG) { | |
4097 | ASSERT(!IO_IS_ALLOCATING(zio)); | |
4098 | checksum = ZIO_CHECKSUM_GANG_HEADER; | |
4099 | } else { | |
4100 | checksum = BP_GET_CHECKSUM(bp); | |
4101 | } | |
4102 | } | |
34dc7c2f | 4103 | |
a6255b7f | 4104 | zio_checksum_compute(zio, checksum, zio->io_abd, zio->io_size); |
34dc7c2f | 4105 | |
62840030 | 4106 | return (zio); |
34dc7c2f BB |
4107 | } |
4108 | ||
62840030 | 4109 | static zio_t * |
b128c09f | 4110 | zio_checksum_verify(zio_t *zio) |
34dc7c2f | 4111 | { |
428870ff | 4112 | zio_bad_cksum_t info; |
b128c09f BB |
4113 | blkptr_t *bp = zio->io_bp; |
4114 | int error; | |
34dc7c2f | 4115 | |
428870ff BB |
4116 | ASSERT(zio->io_vd != NULL); |
4117 | ||
b128c09f BB |
4118 | if (bp == NULL) { |
4119 | /* | |
4120 | * This is zio_read_phys(). | |
4121 | * We're either verifying a label checksum, or nothing at all. | |
4122 | */ | |
4123 | if (zio->io_prop.zp_checksum == ZIO_CHECKSUM_OFF) | |
62840030 | 4124 | return (zio); |
34dc7c2f | 4125 | |
b128c09f BB |
4126 | ASSERT(zio->io_prop.zp_checksum == ZIO_CHECKSUM_LABEL); |
4127 | } | |
34dc7c2f | 4128 | |
428870ff | 4129 | if ((error = zio_checksum_error(zio, &info)) != 0) { |
b128c09f | 4130 | zio->io_error = error; |
7a3066ff MA |
4131 | if (error == ECKSUM && |
4132 | !(zio->io_flags & ZIO_FLAG_SPECULATIVE)) { | |
428870ff | 4133 | zfs_ereport_start_checksum(zio->io_spa, |
b5256303 TC |
4134 | zio->io_vd, &zio->io_bookmark, zio, |
4135 | zio->io_offset, zio->io_size, NULL, &info); | |
b128c09f | 4136 | } |
34dc7c2f BB |
4137 | } |
4138 | ||
62840030 | 4139 | return (zio); |
34dc7c2f BB |
4140 | } |
4141 | ||
4142 | /* | |
4143 | * Called by RAID-Z to ensure we don't compute the checksum twice. | |
4144 | */ | |
4145 | void | |
4146 | zio_checksum_verified(zio_t *zio) | |
4147 | { | |
428870ff | 4148 | zio->io_pipeline &= ~ZIO_STAGE_CHECKSUM_VERIFY; |
34dc7c2f BB |
4149 | } |
4150 | ||
4151 | /* | |
b128c09f BB |
4152 | * ========================================================================== |
4153 | * Error rank. Error are ranked in the order 0, ENXIO, ECKSUM, EIO, other. | |
9b67f605 | 4154 | * An error of 0 indicates success. ENXIO indicates whole-device failure, |
d611989f | 4155 | * which may be transient (e.g. unplugged) or permanent. ECKSUM and EIO |
b128c09f BB |
4156 | * indicate errors that are specific to one I/O, and most likely permanent. |
4157 | * Any other error is presumed to be worse because we weren't expecting it. | |
4158 | * ========================================================================== | |
34dc7c2f | 4159 | */ |
b128c09f BB |
4160 | int |
4161 | zio_worst_error(int e1, int e2) | |
34dc7c2f | 4162 | { |
b128c09f BB |
4163 | static int zio_error_rank[] = { 0, ENXIO, ECKSUM, EIO }; |
4164 | int r1, r2; | |
4165 | ||
4166 | for (r1 = 0; r1 < sizeof (zio_error_rank) / sizeof (int); r1++) | |
4167 | if (e1 == zio_error_rank[r1]) | |
4168 | break; | |
34dc7c2f | 4169 | |
b128c09f BB |
4170 | for (r2 = 0; r2 < sizeof (zio_error_rank) / sizeof (int); r2++) |
4171 | if (e2 == zio_error_rank[r2]) | |
4172 | break; | |
4173 | ||
4174 | return (r1 > r2 ? e1 : e2); | |
34dc7c2f BB |
4175 | } |
4176 | ||
4177 | /* | |
4178 | * ========================================================================== | |
b128c09f | 4179 | * I/O completion |
34dc7c2f BB |
4180 | * ========================================================================== |
4181 | */ | |
62840030 | 4182 | static zio_t * |
b128c09f | 4183 | zio_ready(zio_t *zio) |
34dc7c2f | 4184 | { |
b128c09f | 4185 | blkptr_t *bp = zio->io_bp; |
d164b209 | 4186 | zio_t *pio, *pio_next; |
3dfb57a3 | 4187 | zio_link_t *zl = NULL; |
34dc7c2f | 4188 | |
ddc751d5 GW |
4189 | if (zio_wait_for_children(zio, ZIO_CHILD_GANG_BIT | ZIO_CHILD_DDT_BIT, |
4190 | ZIO_WAIT_READY)) { | |
62840030 | 4191 | return (NULL); |
ddc751d5 | 4192 | } |
34dc7c2f | 4193 | |
9babb374 | 4194 | if (zio->io_ready) { |
b128c09f | 4195 | ASSERT(IO_IS_ALLOCATING(zio)); |
03c6040b GW |
4196 | ASSERT(bp->blk_birth == zio->io_txg || BP_IS_HOLE(bp) || |
4197 | (zio->io_flags & ZIO_FLAG_NOPWRITE)); | |
b128c09f | 4198 | ASSERT(zio->io_children[ZIO_CHILD_GANG][ZIO_WAIT_READY] == 0); |
34dc7c2f | 4199 | |
b128c09f BB |
4200 | zio->io_ready(zio); |
4201 | } | |
34dc7c2f | 4202 | |
b128c09f BB |
4203 | if (bp != NULL && bp != &zio->io_bp_copy) |
4204 | zio->io_bp_copy = *bp; | |
34dc7c2f | 4205 | |
3dfb57a3 | 4206 | if (zio->io_error != 0) { |
b128c09f | 4207 | zio->io_pipeline = ZIO_INTERLOCK_PIPELINE; |
34dc7c2f | 4208 | |
3dfb57a3 DB |
4209 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { |
4210 | ASSERT(IO_IS_ALLOCATING(zio)); | |
4211 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
cc99f275 DB |
4212 | ASSERT(zio->io_metaslab_class != NULL); |
4213 | ||
3dfb57a3 DB |
4214 | /* |
4215 | * We were unable to allocate anything, unreserve and | |
4216 | * issue the next I/O to allocate. | |
4217 | */ | |
4218 | metaslab_class_throttle_unreserve( | |
cc99f275 DB |
4219 | zio->io_metaslab_class, zio->io_prop.zp_copies, |
4220 | zio->io_allocator, zio); | |
492f64e9 | 4221 | zio_allocate_dispatch(zio->io_spa, zio->io_allocator); |
3dfb57a3 DB |
4222 | } |
4223 | } | |
4224 | ||
d164b209 BB |
4225 | mutex_enter(&zio->io_lock); |
4226 | zio->io_state[ZIO_WAIT_READY] = 1; | |
3dfb57a3 | 4227 | pio = zio_walk_parents(zio, &zl); |
d164b209 BB |
4228 | mutex_exit(&zio->io_lock); |
4229 | ||
4230 | /* | |
4231 | * As we notify zio's parents, new parents could be added. | |
4232 | * New parents go to the head of zio's io_parent_list, however, | |
4233 | * so we will (correctly) not notify them. The remainder of zio's | |
4234 | * io_parent_list, from 'pio_next' onward, cannot change because | |
4235 | * all parents must wait for us to be done before they can be done. | |
4236 | */ | |
4237 | for (; pio != NULL; pio = pio_next) { | |
3dfb57a3 | 4238 | pio_next = zio_walk_parents(zio, &zl); |
62840030 | 4239 | zio_notify_parent(pio, zio, ZIO_WAIT_READY, NULL); |
d164b209 | 4240 | } |
34dc7c2f | 4241 | |
428870ff BB |
4242 | if (zio->io_flags & ZIO_FLAG_NODATA) { |
4243 | if (BP_IS_GANG(bp)) { | |
4244 | zio->io_flags &= ~ZIO_FLAG_NODATA; | |
4245 | } else { | |
a6255b7f | 4246 | ASSERT((uintptr_t)zio->io_abd < SPA_MAXBLOCKSIZE); |
428870ff BB |
4247 | zio->io_pipeline &= ~ZIO_VDEV_IO_STAGES; |
4248 | } | |
4249 | } | |
4250 | ||
4251 | if (zio_injection_enabled && | |
4252 | zio->io_spa->spa_syncing_txg == zio->io_txg) | |
4253 | zio_handle_ignored_writes(zio); | |
4254 | ||
62840030 | 4255 | return (zio); |
34dc7c2f BB |
4256 | } |
4257 | ||
3dfb57a3 DB |
4258 | /* |
4259 | * Update the allocation throttle accounting. | |
4260 | */ | |
4261 | static void | |
4262 | zio_dva_throttle_done(zio_t *zio) | |
4263 | { | |
1c27024e | 4264 | ASSERTV(zio_t *lio = zio->io_logical); |
3dfb57a3 DB |
4265 | zio_t *pio = zio_unique_parent(zio); |
4266 | vdev_t *vd = zio->io_vd; | |
4267 | int flags = METASLAB_ASYNC_ALLOC; | |
4268 | ||
4269 | ASSERT3P(zio->io_bp, !=, NULL); | |
4270 | ASSERT3U(zio->io_type, ==, ZIO_TYPE_WRITE); | |
4271 | ASSERT3U(zio->io_priority, ==, ZIO_PRIORITY_ASYNC_WRITE); | |
4272 | ASSERT3U(zio->io_child_type, ==, ZIO_CHILD_VDEV); | |
4273 | ASSERT(vd != NULL); | |
4274 | ASSERT3P(vd, ==, vd->vdev_top); | |
21df134f SB |
4275 | ASSERT(zio_injection_enabled || !(zio->io_flags & ZIO_FLAG_IO_RETRY)); |
4276 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
3dfb57a3 DB |
4277 | ASSERT(zio->io_flags & ZIO_FLAG_IO_ALLOCATING); |
4278 | ASSERT(!(lio->io_flags & ZIO_FLAG_IO_REWRITE)); | |
4279 | ASSERT(!(lio->io_orig_flags & ZIO_FLAG_NODATA)); | |
4280 | ||
4281 | /* | |
4282 | * Parents of gang children can have two flavors -- ones that | |
4283 | * allocated the gang header (will have ZIO_FLAG_IO_REWRITE set) | |
4284 | * and ones that allocated the constituent blocks. The allocation | |
4285 | * throttle needs to know the allocating parent zio so we must find | |
4286 | * it here. | |
4287 | */ | |
4288 | if (pio->io_child_type == ZIO_CHILD_GANG) { | |
4289 | /* | |
4290 | * If our parent is a rewrite gang child then our grandparent | |
4291 | * would have been the one that performed the allocation. | |
4292 | */ | |
4293 | if (pio->io_flags & ZIO_FLAG_IO_REWRITE) | |
4294 | pio = zio_unique_parent(pio); | |
4295 | flags |= METASLAB_GANG_CHILD; | |
4296 | } | |
4297 | ||
4298 | ASSERT(IO_IS_ALLOCATING(pio)); | |
4299 | ASSERT3P(zio, !=, zio->io_logical); | |
4300 | ASSERT(zio->io_logical != NULL); | |
4301 | ASSERT(!(zio->io_flags & ZIO_FLAG_IO_REPAIR)); | |
4302 | ASSERT0(zio->io_flags & ZIO_FLAG_NOPWRITE); | |
cc99f275 | 4303 | ASSERT(zio->io_metaslab_class != NULL); |
3dfb57a3 DB |
4304 | |
4305 | mutex_enter(&pio->io_lock); | |
492f64e9 PD |
4306 | metaslab_group_alloc_decrement(zio->io_spa, vd->vdev_id, pio, flags, |
4307 | pio->io_allocator, B_TRUE); | |
3dfb57a3 DB |
4308 | mutex_exit(&pio->io_lock); |
4309 | ||
cc99f275 DB |
4310 | metaslab_class_throttle_unreserve(zio->io_metaslab_class, 1, |
4311 | pio->io_allocator, pio); | |
3dfb57a3 DB |
4312 | |
4313 | /* | |
4314 | * Call into the pipeline to see if there is more work that | |
4315 | * needs to be done. If there is work to be done it will be | |
4316 | * dispatched to another taskq thread. | |
4317 | */ | |
492f64e9 | 4318 | zio_allocate_dispatch(zio->io_spa, pio->io_allocator); |
3dfb57a3 DB |
4319 | } |
4320 | ||
62840030 | 4321 | static zio_t * |
b128c09f | 4322 | zio_done(zio_t *zio) |
34dc7c2f | 4323 | { |
3dfb57a3 DB |
4324 | /* |
4325 | * Always attempt to keep stack usage minimal here since | |
d611989f | 4326 | * we can be called recursively up to 19 levels deep. |
3dfb57a3 | 4327 | */ |
84c07ada | 4328 | const uint64_t psize = zio->io_size; |
d164b209 | 4329 | zio_t *pio, *pio_next; |
3dfb57a3 | 4330 | zio_link_t *zl = NULL; |
34dc7c2f | 4331 | |
b128c09f | 4332 | /* |
9babb374 | 4333 | * If our children haven't all completed, |
b128c09f BB |
4334 | * wait for them and then repeat this pipeline stage. |
4335 | */ | |
ddc751d5 | 4336 | if (zio_wait_for_children(zio, ZIO_CHILD_ALL_BITS, ZIO_WAIT_DONE)) { |
62840030 | 4337 | return (NULL); |
ddc751d5 | 4338 | } |
34dc7c2f | 4339 | |
3dfb57a3 DB |
4340 | /* |
4341 | * If the allocation throttle is enabled, then update the accounting. | |
4342 | * We only track child I/Os that are part of an allocating async | |
4343 | * write. We must do this since the allocation is performed | |
4344 | * by the logical I/O but the actual write is done by child I/Os. | |
4345 | */ | |
4346 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING && | |
4347 | zio->io_child_type == ZIO_CHILD_VDEV) { | |
cc99f275 DB |
4348 | ASSERT(zio->io_metaslab_class != NULL); |
4349 | ASSERT(zio->io_metaslab_class->mc_alloc_throttle_enabled); | |
3dfb57a3 DB |
4350 | zio_dva_throttle_done(zio); |
4351 | } | |
4352 | ||
4353 | /* | |
4354 | * If the allocation throttle is enabled, verify that | |
4355 | * we have decremented the refcounts for every I/O that was throttled. | |
4356 | */ | |
4357 | if (zio->io_flags & ZIO_FLAG_IO_ALLOCATING) { | |
4358 | ASSERT(zio->io_type == ZIO_TYPE_WRITE); | |
4359 | ASSERT(zio->io_priority == ZIO_PRIORITY_ASYNC_WRITE); | |
4360 | ASSERT(zio->io_bp != NULL); | |
cc99f275 | 4361 | |
492f64e9 PD |
4362 | metaslab_group_alloc_verify(zio->io_spa, zio->io_bp, zio, |
4363 | zio->io_allocator); | |
424fd7c3 | 4364 | VERIFY(zfs_refcount_not_held( |
cc99f275 | 4365 | &zio->io_metaslab_class->mc_alloc_slots[zio->io_allocator], |
492f64e9 | 4366 | zio)); |
3dfb57a3 DB |
4367 | } |
4368 | ||
4369 | ||
1c27024e DB |
4370 | for (int c = 0; c < ZIO_CHILD_TYPES; c++) |
4371 | for (int w = 0; w < ZIO_WAIT_TYPES; w++) | |
b128c09f BB |
4372 | ASSERT(zio->io_children[c][w] == 0); |
4373 | ||
9b67f605 | 4374 | if (zio->io_bp != NULL && !BP_IS_EMBEDDED(zio->io_bp)) { |
c776b317 BB |
4375 | ASSERT(zio->io_bp->blk_pad[0] == 0); |
4376 | ASSERT(zio->io_bp->blk_pad[1] == 0); | |
d1d7e268 MK |
4377 | ASSERT(bcmp(zio->io_bp, &zio->io_bp_copy, |
4378 | sizeof (blkptr_t)) == 0 || | |
c776b317 BB |
4379 | (zio->io_bp == zio_unique_parent(zio)->io_bp)); |
4380 | if (zio->io_type == ZIO_TYPE_WRITE && !BP_IS_HOLE(zio->io_bp) && | |
428870ff | 4381 | zio->io_bp_override == NULL && |
b128c09f | 4382 | !(zio->io_flags & ZIO_FLAG_IO_REPAIR)) { |
d1d7e268 MK |
4383 | ASSERT3U(zio->io_prop.zp_copies, <=, |
4384 | BP_GET_NDVAS(zio->io_bp)); | |
c776b317 | 4385 | ASSERT(BP_COUNT_GANG(zio->io_bp) == 0 || |
d1d7e268 MK |
4386 | (BP_COUNT_GANG(zio->io_bp) == |
4387 | BP_GET_NDVAS(zio->io_bp))); | |
b128c09f | 4388 | } |
03c6040b GW |
4389 | if (zio->io_flags & ZIO_FLAG_NOPWRITE) |
4390 | VERIFY(BP_EQUAL(zio->io_bp, &zio->io_bp_orig)); | |
b128c09f BB |
4391 | } |
4392 | ||
4393 | /* | |
428870ff | 4394 | * If there were child vdev/gang/ddt errors, they apply to us now. |
b128c09f BB |
4395 | */ |
4396 | zio_inherit_child_errors(zio, ZIO_CHILD_VDEV); | |
4397 | zio_inherit_child_errors(zio, ZIO_CHILD_GANG); | |
428870ff BB |
4398 | zio_inherit_child_errors(zio, ZIO_CHILD_DDT); |
4399 | ||
4400 | /* | |
4401 | * If the I/O on the transformed data was successful, generate any | |
4402 | * checksum reports now while we still have the transformed data. | |
4403 | */ | |
4404 | if (zio->io_error == 0) { | |
4405 | while (zio->io_cksum_report != NULL) { | |
4406 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4407 | uint64_t align = zcr->zcr_align; | |
a6255b7f | 4408 | uint64_t asize = P2ROUNDUP(psize, align); |
a6255b7f DQ |
4409 | abd_t *adata = zio->io_abd; |
4410 | ||
4411 | if (asize != psize) { | |
84c07ada | 4412 | adata = abd_alloc(asize, B_TRUE); |
a6255b7f DQ |
4413 | abd_copy(adata, zio->io_abd, psize); |
4414 | abd_zero_off(adata, psize, asize - psize); | |
428870ff BB |
4415 | } |
4416 | ||
4417 | zio->io_cksum_report = zcr->zcr_next; | |
4418 | zcr->zcr_next = NULL; | |
84c07ada | 4419 | zcr->zcr_finish(zcr, adata); |
428870ff BB |
4420 | zfs_ereport_free_checksum(zcr); |
4421 | ||
a6255b7f DQ |
4422 | if (asize != psize) |
4423 | abd_free(adata); | |
428870ff BB |
4424 | } |
4425 | } | |
b128c09f BB |
4426 | |
4427 | zio_pop_transforms(zio); /* note: may set zio->io_error */ | |
4428 | ||
a6255b7f | 4429 | vdev_stat_update(zio, psize); |
b128c09f | 4430 | |
a69052be | 4431 | /* |
cc92e9d0 | 4432 | * If this I/O is attached to a particular vdev is slow, exceeding |
72f53c56 MJ |
4433 | * 30 seconds to complete, post an error described the I/O delay. |
4434 | * We ignore these errors if the device is currently unavailable. | |
a69052be | 4435 | */ |
ad796b8a TH |
4436 | if (zio->io_delay >= MSEC2NSEC(zio_slow_io_ms)) { |
4437 | if (zio->io_vd != NULL && !vdev_is_dead(zio->io_vd)) { | |
4438 | /* | |
4439 | * We want to only increment our slow IO counters if | |
4440 | * the IO is valid (i.e. not if the drive is removed). | |
4441 | * | |
4442 | * zfs_ereport_post() will also do these checks, but | |
4443 | * it can also ratelimit and have other failures, so we | |
4444 | * need to increment the slow_io counters independent | |
4445 | * of it. | |
4446 | */ | |
4447 | if (zfs_ereport_is_valid(FM_EREPORT_ZFS_DELAY, | |
4448 | zio->io_spa, zio->io_vd, zio)) { | |
4449 | mutex_enter(&zio->io_vd->vdev_stat_lock); | |
4450 | zio->io_vd->vdev_stat.vs_slow_ios++; | |
4451 | mutex_exit(&zio->io_vd->vdev_stat_lock); | |
4452 | ||
4453 | zfs_ereport_post(FM_EREPORT_ZFS_DELAY, | |
4454 | zio->io_spa, zio->io_vd, &zio->io_bookmark, | |
4455 | zio, 0, 0); | |
4456 | } | |
4457 | } | |
72f53c56 | 4458 | } |
a69052be | 4459 | |
b128c09f BB |
4460 | if (zio->io_error) { |
4461 | /* | |
4462 | * If this I/O is attached to a particular vdev, | |
4463 | * generate an error message describing the I/O failure | |
4464 | * at the block level. We ignore these errors if the | |
4465 | * device is currently unavailable. | |
4466 | */ | |
c776b317 | 4467 | if (zio->io_error != ECKSUM && zio->io_vd != NULL && |
02730c33 | 4468 | !vdev_is_dead(zio->io_vd)) |
c776b317 | 4469 | zfs_ereport_post(FM_EREPORT_ZFS_IO, zio->io_spa, |
b5256303 | 4470 | zio->io_vd, &zio->io_bookmark, zio, 0, 0); |
34dc7c2f | 4471 | |
428870ff BB |
4472 | if ((zio->io_error == EIO || !(zio->io_flags & |
4473 | (ZIO_FLAG_SPECULATIVE | ZIO_FLAG_DONT_PROPAGATE))) && | |
c776b317 | 4474 | zio == zio->io_logical) { |
b128c09f BB |
4475 | /* |
4476 | * For logical I/O requests, tell the SPA to log the | |
4477 | * error and generate a logical data ereport. | |
4478 | */ | |
b5256303 | 4479 | spa_log_error(zio->io_spa, &zio->io_bookmark); |
d1d7e268 | 4480 | zfs_ereport_post(FM_EREPORT_ZFS_DATA, zio->io_spa, |
b5256303 | 4481 | NULL, &zio->io_bookmark, zio, 0, 0); |
b128c09f BB |
4482 | } |
4483 | } | |
34dc7c2f | 4484 | |
c776b317 | 4485 | if (zio->io_error && zio == zio->io_logical) { |
b128c09f BB |
4486 | /* |
4487 | * Determine whether zio should be reexecuted. This will | |
4488 | * propagate all the way to the root via zio_notify_parent(). | |
4489 | */ | |
c776b317 | 4490 | ASSERT(zio->io_vd == NULL && zio->io_bp != NULL); |
428870ff | 4491 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); |
b128c09f | 4492 | |
428870ff BB |
4493 | if (IO_IS_ALLOCATING(zio) && |
4494 | !(zio->io_flags & ZIO_FLAG_CANFAIL)) { | |
b128c09f BB |
4495 | if (zio->io_error != ENOSPC) |
4496 | zio->io_reexecute |= ZIO_REEXECUTE_NOW; | |
4497 | else | |
4498 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff | 4499 | } |
b128c09f BB |
4500 | |
4501 | if ((zio->io_type == ZIO_TYPE_READ || | |
4502 | zio->io_type == ZIO_TYPE_FREE) && | |
572e2857 | 4503 | !(zio->io_flags & ZIO_FLAG_SCAN_THREAD) && |
b128c09f | 4504 | zio->io_error == ENXIO && |
c776b317 BB |
4505 | spa_load_state(zio->io_spa) == SPA_LOAD_NONE && |
4506 | spa_get_failmode(zio->io_spa) != ZIO_FAILURE_MODE_CONTINUE) | |
b128c09f BB |
4507 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; |
4508 | ||
4509 | if (!(zio->io_flags & ZIO_FLAG_CANFAIL) && !zio->io_reexecute) | |
4510 | zio->io_reexecute |= ZIO_REEXECUTE_SUSPEND; | |
428870ff BB |
4511 | |
4512 | /* | |
4513 | * Here is a possibly good place to attempt to do | |
4514 | * either combinatorial reconstruction or error correction | |
4515 | * based on checksums. It also might be a good place | |
4516 | * to send out preliminary ereports before we suspend | |
4517 | * processing. | |
4518 | */ | |
34dc7c2f BB |
4519 | } |
4520 | ||
4521 | /* | |
b128c09f BB |
4522 | * If there were logical child errors, they apply to us now. |
4523 | * We defer this until now to avoid conflating logical child | |
4524 | * errors with errors that happened to the zio itself when | |
4525 | * updating vdev stats and reporting FMA events above. | |
34dc7c2f | 4526 | */ |
b128c09f | 4527 | zio_inherit_child_errors(zio, ZIO_CHILD_LOGICAL); |
34dc7c2f | 4528 | |
428870ff BB |
4529 | if ((zio->io_error || zio->io_reexecute) && |
4530 | IO_IS_ALLOCATING(zio) && zio->io_gang_leader == zio && | |
03c6040b | 4531 | !(zio->io_flags & (ZIO_FLAG_IO_REWRITE | ZIO_FLAG_NOPWRITE))) |
c776b317 | 4532 | zio_dva_unallocate(zio, zio->io_gang_tree, zio->io_bp); |
9babb374 BB |
4533 | |
4534 | zio_gang_tree_free(&zio->io_gang_tree); | |
4535 | ||
4536 | /* | |
4537 | * Godfather I/Os should never suspend. | |
4538 | */ | |
4539 | if ((zio->io_flags & ZIO_FLAG_GODFATHER) && | |
4540 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) | |
a32494d2 | 4541 | zio->io_reexecute &= ~ZIO_REEXECUTE_SUSPEND; |
9babb374 | 4542 | |
b128c09f BB |
4543 | if (zio->io_reexecute) { |
4544 | /* | |
4545 | * This is a logical I/O that wants to reexecute. | |
4546 | * | |
4547 | * Reexecute is top-down. When an i/o fails, if it's not | |
4548 | * the root, it simply notifies its parent and sticks around. | |
4549 | * The parent, seeing that it still has children in zio_done(), | |
4550 | * does the same. This percolates all the way up to the root. | |
4551 | * The root i/o will reexecute or suspend the entire tree. | |
4552 | * | |
4553 | * This approach ensures that zio_reexecute() honors | |
4554 | * all the original i/o dependency relationships, e.g. | |
4555 | * parents not executing until children are ready. | |
4556 | */ | |
4557 | ASSERT(zio->io_child_type == ZIO_CHILD_LOGICAL); | |
34dc7c2f | 4558 | |
9babb374 | 4559 | zio->io_gang_leader = NULL; |
b128c09f | 4560 | |
d164b209 BB |
4561 | mutex_enter(&zio->io_lock); |
4562 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4563 | mutex_exit(&zio->io_lock); | |
4564 | ||
9babb374 BB |
4565 | /* |
4566 | * "The Godfather" I/O monitors its children but is | |
4567 | * not a true parent to them. It will track them through | |
4568 | * the pipeline but severs its ties whenever they get into | |
4569 | * trouble (e.g. suspended). This allows "The Godfather" | |
4570 | * I/O to return status without blocking. | |
4571 | */ | |
3dfb57a3 DB |
4572 | zl = NULL; |
4573 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; | |
4574 | pio = pio_next) { | |
4575 | zio_link_t *remove_zl = zl; | |
4576 | pio_next = zio_walk_parents(zio, &zl); | |
9babb374 BB |
4577 | |
4578 | if ((pio->io_flags & ZIO_FLAG_GODFATHER) && | |
4579 | (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND)) { | |
3dfb57a3 | 4580 | zio_remove_child(pio, zio, remove_zl); |
62840030 MA |
4581 | /* |
4582 | * This is a rare code path, so we don't | |
4583 | * bother with "next_to_execute". | |
4584 | */ | |
4585 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, | |
4586 | NULL); | |
9babb374 BB |
4587 | } |
4588 | } | |
4589 | ||
d164b209 | 4590 | if ((pio = zio_unique_parent(zio)) != NULL) { |
b128c09f BB |
4591 | /* |
4592 | * We're not a root i/o, so there's nothing to do | |
4593 | * but notify our parent. Don't propagate errors | |
4594 | * upward since we haven't permanently failed yet. | |
4595 | */ | |
9babb374 | 4596 | ASSERT(!(zio->io_flags & ZIO_FLAG_GODFATHER)); |
b128c09f | 4597 | zio->io_flags |= ZIO_FLAG_DONT_PROPAGATE; |
62840030 MA |
4598 | /* |
4599 | * This is a rare code path, so we don't bother with | |
4600 | * "next_to_execute". | |
4601 | */ | |
4602 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, NULL); | |
b128c09f BB |
4603 | } else if (zio->io_reexecute & ZIO_REEXECUTE_SUSPEND) { |
4604 | /* | |
4605 | * We'd fail again if we reexecuted now, so suspend | |
4606 | * until conditions improve (e.g. device comes online). | |
4607 | */ | |
cec3a0a1 | 4608 | zio_suspend(zio->io_spa, zio, ZIO_SUSPEND_IOERR); |
b128c09f BB |
4609 | } else { |
4610 | /* | |
4611 | * Reexecution is potentially a huge amount of work. | |
4612 | * Hand it off to the otherwise-unused claim taskq. | |
4613 | */ | |
a38718a6 | 4614 | ASSERT(taskq_empty_ent(&zio->io_tqent)); |
7ef5e54e AL |
4615 | spa_taskq_dispatch_ent(zio->io_spa, |
4616 | ZIO_TYPE_CLAIM, ZIO_TASKQ_ISSUE, | |
a38718a6 GA |
4617 | (task_func_t *)zio_reexecute, zio, 0, |
4618 | &zio->io_tqent); | |
b128c09f | 4619 | } |
62840030 | 4620 | return (NULL); |
34dc7c2f BB |
4621 | } |
4622 | ||
428870ff | 4623 | ASSERT(zio->io_child_count == 0); |
b128c09f BB |
4624 | ASSERT(zio->io_reexecute == 0); |
4625 | ASSERT(zio->io_error == 0 || (zio->io_flags & ZIO_FLAG_CANFAIL)); | |
34dc7c2f | 4626 | |
428870ff BB |
4627 | /* |
4628 | * Report any checksum errors, since the I/O is complete. | |
4629 | */ | |
4630 | while (zio->io_cksum_report != NULL) { | |
4631 | zio_cksum_report_t *zcr = zio->io_cksum_report; | |
4632 | zio->io_cksum_report = zcr->zcr_next; | |
4633 | zcr->zcr_next = NULL; | |
4634 | zcr->zcr_finish(zcr, NULL); | |
4635 | zfs_ereport_free_checksum(zcr); | |
4636 | } | |
4637 | ||
920dd524 | 4638 | if (zio->io_flags & ZIO_FLAG_FASTWRITE && zio->io_bp && |
9b67f605 MA |
4639 | !BP_IS_HOLE(zio->io_bp) && !BP_IS_EMBEDDED(zio->io_bp) && |
4640 | !(zio->io_flags & ZIO_FLAG_NOPWRITE)) { | |
920dd524 ED |
4641 | metaslab_fastwrite_unmark(zio->io_spa, zio->io_bp); |
4642 | } | |
4643 | ||
d164b209 BB |
4644 | /* |
4645 | * It is the responsibility of the done callback to ensure that this | |
4646 | * particular zio is no longer discoverable for adoption, and as | |
4647 | * such, cannot acquire any new parents. | |
4648 | */ | |
b128c09f BB |
4649 | if (zio->io_done) |
4650 | zio->io_done(zio); | |
34dc7c2f | 4651 | |
d164b209 BB |
4652 | mutex_enter(&zio->io_lock); |
4653 | zio->io_state[ZIO_WAIT_DONE] = 1; | |
4654 | mutex_exit(&zio->io_lock); | |
34dc7c2f | 4655 | |
62840030 MA |
4656 | /* |
4657 | * We are done executing this zio. We may want to execute a parent | |
4658 | * next. See the comment in zio_notify_parent(). | |
4659 | */ | |
4660 | zio_t *next_to_execute = NULL; | |
3dfb57a3 DB |
4661 | zl = NULL; |
4662 | for (pio = zio_walk_parents(zio, &zl); pio != NULL; pio = pio_next) { | |
4663 | zio_link_t *remove_zl = zl; | |
4664 | pio_next = zio_walk_parents(zio, &zl); | |
4665 | zio_remove_child(pio, zio, remove_zl); | |
62840030 | 4666 | zio_notify_parent(pio, zio, ZIO_WAIT_DONE, &next_to_execute); |
b128c09f | 4667 | } |
34dc7c2f | 4668 | |
b128c09f BB |
4669 | if (zio->io_waiter != NULL) { |
4670 | mutex_enter(&zio->io_lock); | |
4671 | zio->io_executor = NULL; | |
4672 | cv_broadcast(&zio->io_cv); | |
4673 | mutex_exit(&zio->io_lock); | |
4674 | } else { | |
4675 | zio_destroy(zio); | |
4676 | } | |
34dc7c2f | 4677 | |
62840030 | 4678 | return (next_to_execute); |
34dc7c2f BB |
4679 | } |
4680 | ||
4681 | /* | |
b128c09f BB |
4682 | * ========================================================================== |
4683 | * I/O pipeline definition | |
4684 | * ========================================================================== | |
34dc7c2f | 4685 | */ |
428870ff | 4686 | static zio_pipe_stage_t *zio_pipeline[] = { |
b128c09f | 4687 | NULL, |
b128c09f | 4688 | zio_read_bp_init, |
3dfb57a3 | 4689 | zio_write_bp_init, |
428870ff BB |
4690 | zio_free_bp_init, |
4691 | zio_issue_async, | |
3dfb57a3 | 4692 | zio_write_compress, |
b5256303 | 4693 | zio_encrypt, |
b128c09f | 4694 | zio_checksum_generate, |
03c6040b | 4695 | zio_nop_write, |
428870ff BB |
4696 | zio_ddt_read_start, |
4697 | zio_ddt_read_done, | |
4698 | zio_ddt_write, | |
4699 | zio_ddt_free, | |
b128c09f BB |
4700 | zio_gang_assemble, |
4701 | zio_gang_issue, | |
3dfb57a3 | 4702 | zio_dva_throttle, |
b128c09f BB |
4703 | zio_dva_allocate, |
4704 | zio_dva_free, | |
4705 | zio_dva_claim, | |
4706 | zio_ready, | |
4707 | zio_vdev_io_start, | |
4708 | zio_vdev_io_done, | |
4709 | zio_vdev_io_assess, | |
4710 | zio_checksum_verify, | |
4711 | zio_done | |
4712 | }; | |
c28b2279 | 4713 | |
9ae529ec | 4714 | |
9ae529ec | 4715 | |
9ae529ec | 4716 | |
fcff0f35 PD |
4717 | /* |
4718 | * Compare two zbookmark_phys_t's to see which we would reach first in a | |
4719 | * pre-order traversal of the object tree. | |
4720 | * | |
4721 | * This is simple in every case aside from the meta-dnode object. For all other | |
4722 | * objects, we traverse them in order (object 1 before object 2, and so on). | |
4723 | * However, all of these objects are traversed while traversing object 0, since | |
4724 | * the data it points to is the list of objects. Thus, we need to convert to a | |
4725 | * canonical representation so we can compare meta-dnode bookmarks to | |
4726 | * non-meta-dnode bookmarks. | |
4727 | * | |
4728 | * We do this by calculating "equivalents" for each field of the zbookmark. | |
4729 | * zbookmarks outside of the meta-dnode use their own object and level, and | |
4730 | * calculate the level 0 equivalent (the first L0 blkid that is contained in the | |
4731 | * blocks this bookmark refers to) by multiplying their blkid by their span | |
4732 | * (the number of L0 blocks contained within one block at their level). | |
4733 | * zbookmarks inside the meta-dnode calculate their object equivalent | |
4734 | * (which is L0equiv * dnodes per data block), use 0 for their L0equiv, and use | |
4735 | * level + 1<<31 (any value larger than a level could ever be) for their level. | |
4736 | * This causes them to always compare before a bookmark in their object | |
4737 | * equivalent, compare appropriately to bookmarks in other objects, and to | |
4738 | * compare appropriately to other bookmarks in the meta-dnode. | |
4739 | */ | |
4740 | int | |
4741 | zbookmark_compare(uint16_t dbss1, uint8_t ibs1, uint16_t dbss2, uint8_t ibs2, | |
4742 | const zbookmark_phys_t *zb1, const zbookmark_phys_t *zb2) | |
4743 | { | |
4744 | /* | |
4745 | * These variables represent the "equivalent" values for the zbookmark, | |
4746 | * after converting zbookmarks inside the meta dnode to their | |
4747 | * normal-object equivalents. | |
4748 | */ | |
4749 | uint64_t zb1obj, zb2obj; | |
4750 | uint64_t zb1L0, zb2L0; | |
4751 | uint64_t zb1level, zb2level; | |
4752 | ||
4753 | if (zb1->zb_object == zb2->zb_object && | |
4754 | zb1->zb_level == zb2->zb_level && | |
4755 | zb1->zb_blkid == zb2->zb_blkid) | |
4756 | return (0); | |
9ae529ec | 4757 | |
fcff0f35 PD |
4758 | /* |
4759 | * BP_SPANB calculates the span in blocks. | |
4760 | */ | |
4761 | zb1L0 = (zb1->zb_blkid) * BP_SPANB(ibs1, zb1->zb_level); | |
4762 | zb2L0 = (zb2->zb_blkid) * BP_SPANB(ibs2, zb2->zb_level); | |
9ae529ec CS |
4763 | |
4764 | if (zb1->zb_object == DMU_META_DNODE_OBJECT) { | |
fcff0f35 PD |
4765 | zb1obj = zb1L0 * (dbss1 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); |
4766 | zb1L0 = 0; | |
4767 | zb1level = zb1->zb_level + COMPARE_META_LEVEL; | |
4768 | } else { | |
4769 | zb1obj = zb1->zb_object; | |
4770 | zb1level = zb1->zb_level; | |
9ae529ec CS |
4771 | } |
4772 | ||
fcff0f35 PD |
4773 | if (zb2->zb_object == DMU_META_DNODE_OBJECT) { |
4774 | zb2obj = zb2L0 * (dbss2 << (SPA_MINBLOCKSHIFT - DNODE_SHIFT)); | |
4775 | zb2L0 = 0; | |
4776 | zb2level = zb2->zb_level + COMPARE_META_LEVEL; | |
4777 | } else { | |
4778 | zb2obj = zb2->zb_object; | |
4779 | zb2level = zb2->zb_level; | |
4780 | } | |
4781 | ||
4782 | /* Now that we have a canonical representation, do the comparison. */ | |
4783 | if (zb1obj != zb2obj) | |
4784 | return (zb1obj < zb2obj ? -1 : 1); | |
4785 | else if (zb1L0 != zb2L0) | |
4786 | return (zb1L0 < zb2L0 ? -1 : 1); | |
4787 | else if (zb1level != zb2level) | |
4788 | return (zb1level > zb2level ? -1 : 1); | |
4789 | /* | |
4790 | * This can (theoretically) happen if the bookmarks have the same object | |
4791 | * and level, but different blkids, if the block sizes are not the same. | |
4792 | * There is presently no way to change the indirect block sizes | |
4793 | */ | |
4794 | return (0); | |
4795 | } | |
4796 | ||
4797 | /* | |
4798 | * This function checks the following: given that last_block is the place that | |
4799 | * our traversal stopped last time, does that guarantee that we've visited | |
4800 | * every node under subtree_root? Therefore, we can't just use the raw output | |
4801 | * of zbookmark_compare. We have to pass in a modified version of | |
4802 | * subtree_root; by incrementing the block id, and then checking whether | |
4803 | * last_block is before or equal to that, we can tell whether or not having | |
4804 | * visited last_block implies that all of subtree_root's children have been | |
4805 | * visited. | |
4806 | */ | |
4807 | boolean_t | |
4808 | zbookmark_subtree_completed(const dnode_phys_t *dnp, | |
4809 | const zbookmark_phys_t *subtree_root, const zbookmark_phys_t *last_block) | |
4810 | { | |
4811 | zbookmark_phys_t mod_zb = *subtree_root; | |
4812 | mod_zb.zb_blkid++; | |
4813 | ASSERT(last_block->zb_level == 0); | |
4814 | ||
4815 | /* The objset_phys_t isn't before anything. */ | |
4816 | if (dnp == NULL) | |
9ae529ec | 4817 | return (B_FALSE); |
fcff0f35 PD |
4818 | |
4819 | /* | |
4820 | * We pass in 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT) for the | |
4821 | * data block size in sectors, because that variable is only used if | |
4822 | * the bookmark refers to a block in the meta-dnode. Since we don't | |
4823 | * know without examining it what object it refers to, and there's no | |
4824 | * harm in passing in this value in other cases, we always pass it in. | |
4825 | * | |
4826 | * We pass in 0 for the indirect block size shift because zb2 must be | |
4827 | * level 0. The indirect block size is only used to calculate the span | |
4828 | * of the bookmark, but since the bookmark must be level 0, the span is | |
4829 | * always 1, so the math works out. | |
4830 | * | |
4831 | * If you make changes to how the zbookmark_compare code works, be sure | |
4832 | * to make sure that this code still works afterwards. | |
4833 | */ | |
4834 | return (zbookmark_compare(dnp->dn_datablkszsec, dnp->dn_indblkshift, | |
4835 | 1ULL << (DNODE_BLOCK_SHIFT - SPA_MINBLOCKSHIFT), 0, &mod_zb, | |
4836 | last_block) <= 0); | |
9ae529ec CS |
4837 | } |
4838 | ||
93ce2b4c | 4839 | #if defined(_KERNEL) |
c28b2279 | 4840 | EXPORT_SYMBOL(zio_type_name); |
81971b13 BB |
4841 | EXPORT_SYMBOL(zio_buf_alloc); |
4842 | EXPORT_SYMBOL(zio_data_buf_alloc); | |
4843 | EXPORT_SYMBOL(zio_buf_free); | |
4844 | EXPORT_SYMBOL(zio_data_buf_free); | |
c28b2279 | 4845 | |
ad796b8a TH |
4846 | module_param(zio_slow_io_ms, int, 0644); |
4847 | MODULE_PARM_DESC(zio_slow_io_ms, | |
4848 | "Max I/O completion time (milliseconds) before marking it as slow"); | |
c409e464 BB |
4849 | |
4850 | module_param(zio_requeue_io_start_cut_in_line, int, 0644); | |
4851 | MODULE_PARM_DESC(zio_requeue_io_start_cut_in_line, "Prioritize requeued I/O"); | |
29dee3ee CP |
4852 | |
4853 | module_param(zfs_sync_pass_deferred_free, int, 0644); | |
4854 | MODULE_PARM_DESC(zfs_sync_pass_deferred_free, | |
d1d7e268 | 4855 | "Defer frees starting in this pass"); |
29dee3ee CP |
4856 | |
4857 | module_param(zfs_sync_pass_dont_compress, int, 0644); | |
4858 | MODULE_PARM_DESC(zfs_sync_pass_dont_compress, | |
d1d7e268 | 4859 | "Don't compress starting in this pass"); |
29dee3ee CP |
4860 | |
4861 | module_param(zfs_sync_pass_rewrite, int, 0644); | |
4862 | MODULE_PARM_DESC(zfs_sync_pass_rewrite, | |
d1d7e268 | 4863 | "Rewrite new bps starting in this pass"); |
3dfb57a3 DB |
4864 | |
4865 | module_param(zio_dva_throttle_enabled, int, 0644); | |
4866 | MODULE_PARM_DESC(zio_dva_throttle_enabled, | |
4867 | "Throttle block allocations in the ZIO pipeline"); | |
c28b2279 | 4868 | #endif |